feat: release guided Go2 standing skill with explicit drift waiver

Add loaded-standing physics, PPO training, fixed-candidate acceptance, isolated inference, and fail-closed publication. Preserve the original 120-second drift failure and bind the operator-approved exception to the exact candidate and evidence.

Include pending Reach training/release infrastructure and archived packages. Leave pre-existing short-worker visualization changes and large generated training artifacts out of this commit.
This commit is contained in:
2026-10-12 17:26:03 +08:00
parent 8019d6f8d6
commit 093ab883be
145 changed files with 52808 additions and 6 deletions
+11
View File
@@ -0,0 +1,11 @@
成熟机器人资源在:/home/cen/Embodied_Workspace/unitree_ros/robots
你可以使用Isaac Sim mcp直接操作软件
项目计划参考'/home/cen/Documents/Platform'下的文档
第一阶段先考虑四足狗的腿
可以调用子agent加速开发
当前首要任务:Go2 FL受载站立,见 docs/loaded_stand_training.md 和 docs/loaded_stand_acceptance.md。
首轮2048更新已完成,1024候选holdout/full通过,120秒耐久足端漂移30.78mm>30mm,原始验收失败。
用户随后明确同意此微小偏差特例放行,已完成 go2_fl_loaded_stand-1.0.0 带豁免冻结发布,见 docs/loaded_stand_release_1_0_0.md;仅当前候选/证据的foot_drift豁免,不是全局阈值改变。
保留固定候选和原始失败证据;不得以已查看holdout重选权重、缩短耐久或豁免其它门槛。新一轮训练需用户新授权。
旧固定悬空Reach 1.0.1保持归档不变;不以站立控制基线、短训练或奖励上涨替代新策略验收。
+26 -2
View File
@@ -2,7 +2,31 @@
仓库:[chenlin/SkillFactory](https://gitea.robotquan.com/chenlin/SkillFactory)。
真实单腿资产的导入、固定支座仿真与共享策略训练原型。当前覆盖 **Go2 FL / Go1 FL**;第一周 Go2 基准已完成验收,共享 PPO 仅通过短更新测试,**尚无已收敛、冻结或验证泛化的 Skill**。
真实单腿资产导入、固定悬空/可运动承载底座仿真与策略训练。资产/旧接口覆盖 **Go2 FL / Go1 FL**;新承载站立试验仅限Go2 FL,共享 PPO 仍仅为短更新原型。第二阶段新增 **Go2 FL 固定悬空足端定点到达**的固定数据 IQL 训练、恢复、独立评估与冻结包链路;发布验收结果见下文,不声明 Go1/整机/接触迁移。
## 当前首要任务:受载站立策略训练
主线已切换到 `SkillFactory-LoadedLegStand-v0` 的独立PPO训练。使用真实六维测力、可运动3kg底座和地面足端接触;保留横向/转动导向,竖直辅助为0。**旧Reach冻结权重不用于新任务。**
- [训练预登记、预算与选模规则](docs/loaded_stand_training.md)
- 配置:`configs/training/go2_fl_loaded_stand_ppo_v1.json`
- 进度:`logs/loaded_stand/train_v1/progress.json`;检查点:该目录的 `checkpoints/`
- **首轮已完成**:256环境、2048次更新、16,777,216条转移,耗时25.18分钟。validation-only选中1024更新,平均高度误差0.9663mm;随后holdout均值0.921mm、13类全幅通过,但原始120秒耐久足端漂移30.78mm超过30mm。**用户批准本次固定偏差后,已带豁免冻结发布1.0.0**;不是原始门槛全部通过。
- 候选:`logs/loaded_stand/train_v1/checkpoints/best_validation.pt`;[原始验收报告](docs/loaded_stand_acceptance.md);[用户豁免与1.0.0发布报告](docs/loaded_stand_release_1_0_0.md)。归档:`releases/go2_fl_loaded_stand-1.0.0.tar.gz`。原始失败证据保留,不以已查看holdout重选权重。
## 已完成归档:Reach离线预训练与冻结包
- 范围与预注册门槛:[`docs/stage2_plan.md`](docs/stage2_plan.md);复现/推理:[`docs/stage2_usage.md`](docs/stage2_usage.md)。
- 最新验收报告:[`docs/stage2_release_1_0_1.md`](docs/stage2_release_1_0_1.md);正式训练配置:`configs/training/go2_fl_offline_v1_attempt2.json`。
- 发布包:[`releases/go2_fl_reach-1.0.1.tar.gz`](releases/go2_fl_reach-1.0.1.tar.gz);只读目录:`releases/go2_fl_reach-1.0.1/`。完整边界6面/12边/8角物理复验通过(均值2.308mm),统一12项验收全部通过,权重未变。独立SkillBundle v1/leg-token-v1,不是通用SoftwareSpec v2 adapter。
- 新复验证据:`logs/stage2/v1_0_1/`;原训练/恢复点:`logs/stage2/v1_attempt2/`;失败首轮:`logs/stage2/v1/`。1.0.0及[历史报告](docs/stage2_acceptance.md)原样保留,其旧`accepted`不代表新门槛通过,见[修复记录](docs/stage2_reliability_fix.md)。
## 新任务:连接底座的受载站立(实验v0)
- `SkillFactory-LoadedLegStand-v0`:3kg可运动虚拟底座、刚性安装六维测力、足端接地;横向/转动弹簧阻尼导向,**竖直导向力为0**,没有世界固定根。
- 13种无扰动/六轴正负载荷场景×14秒真实物理验证通过;最坏高度误差8.01mm。关闭驱动后全部环境向下塌落约35.46cm,排除了竖直悬挂承重。
- **本节数据属于反馈控制基线和训练环境;独立RL首轮已另行训练并获有限漂移豁免发布,但不是原始门槛全通过或无辅助单足/整机平衡。** 新观测含6D安装口wrench,与Reach冻结接口不兼容。
- [模型、六维接口、实测报告与启动命令](docs/loaded_stand.md);配置:`configs/standing/go2_fl_support_v0.json`;证据:`logs/loaded_stand/v0/`。旧Reach资产和1.0.1包不变。
## 统一软硬件描述规范
@@ -77,4 +101,4 @@ PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 ANCHOR_RUNTIME_TESTS=1 "$LAB" -p -m pytest test
- `SkillFactory-SingleLegAnchor-Go2-v0`:保留 `[B,19]` 验收接口。
- `Template-Skillfactory-v0`:原 Cartpole 与 `scripts/rsl_rl/` 入口保留。
生成资产、日志、checkpoint 不入库;换机器需重建复验。当前采样器每轮重启 Kit,仅适合短接口测试,**不要据此直接长训或宣称高吞吐**。后续优先常驻采样、留出形态评估,再实现 latent 学习和冻结调用;固定支座结果不证明整机迁移。
生成资产、训练日志、checkpoint 不入库;换机器需重建复验,冻结发布包独立保留。旧 `train_leg_skill.py` 每轮重启 Kit,仍仅适合短接口测试;正式离线链路使用常驻收集器与独立无仿真梯度训练。后续 latent 学习、留出形态评估和整机协调仍未实现;固定支座结果不证明整机迁移。
+28
View File
@@ -0,0 +1,28 @@
{
"schema_version": 1,
"task": "go2_fl_loaded_stand_v0",
"source_bundle": "assets/generated/legs/go2_fl",
"support_asset": "assets/generated/support/go2_fl_3kg_v0",
"payload_mass_kg": 3.0,
"payload_size_m": [0.18, 0.18, 0.08],
"payload_com_m": [0.0, 0.0, 0.06],
"sensor_carrier_mass_kg": 0.05,
"guide_stiffness": [250.0, 250.0, 0.0, 100.0, 100.0, 100.0],
"guide_damping": [20.0, 20.0, 0.0, 5.0, 5.0, 5.0],
"guide_wrench_limit": [60.0, 60.0, 0.0, 25.0, 25.0, 25.0],
"disturbance_amplitude": [5.0, 5.0, 10.0, 1.0, 1.0, 1.0],
"action_scale_rad": 0.5,
"initial_foot_clearance_m": 0.003,
"episode_seconds": 14.0,
"warmup_seconds": 2.0,
"height_error_limit_m": 0.06,
"tilt_limit_rad": 0.2617993877991494,
"foot_drift_limit_m": 0.03,
"contact_fraction_min": 0.98,
"vertical_balance_relative_tolerance": 0.10,
"sensor_static_force_tolerance_N": 1.5,
"sensor_static_torque_tolerance_Nm": 0.3,
"normalized_effort_velocity_limit": 1.00001,
"joint_limit_excess_rad": 0.01,
"seed": 71000
}
@@ -0,0 +1,30 @@
{
"format": "skillfactory.loaded-stand-ppo.v1",
"task_config": "configs/standing/go2_fl_support_v0.json",
"seed": 81000,
"validation_seed": 82000,
"holdout_seed": 83000,
"num_envs": 256,
"horizon": 32,
"updates": 2048,
"hidden": 128,
"initial_std": 0.15,
"learning_rate": 0.0003,
"gamma": 0.995,
"gae_lambda": 0.95,
"clip_ratio": 0.2,
"epochs": 4,
"minibatch_size": 2048,
"value_coefficient": 0.5,
"max_gradient_norm": 1.0,
"target_kl": 0.02,
"failure_penalty": 5.0,
"checkpoint_interval": 128,
"validation_interval": 256,
"train_amplitude_scale": [0.5, 1.0],
"train_time_shift_s": [-0.5, 0.5],
"validation_amplitude_scale": [0.6, 1.0],
"validation_time_shift_s": [-0.25, 0.25],
"max_wall_seconds": 5400,
"max_disk_bytes": 2147483648
}
+17
View File
@@ -0,0 +1,17 @@
{
"experiment": "go2_fl_reach_iql_v1",
"scope": "Go2 FL only; fixed suspended anchor; 3-DoF foot-link Cartesian stationary reach; nominal reset; no contact, noise, morphology or whole-body transfer claims",
"bundle": "assets/generated/legs/go2_fl",
"max_joints": 8,
"num_envs": 256,
"episode_steps": 200,
"settle_steps": 150,
"goal_action_range": 0.75,
"collection": {"seed": 11000, "shards": 20, "noise_std_by_quartile": [0.02, 0.08, 0.2, 0.4], "feedback_gain": 1.5, "velocity_gain_s": 0.04},
"training": {"seed": 42, "updates": 30000, "batch_size": 512, "hidden": 128, "lr": 0.0003, "discount": 0.99, "expectile": 0.7, "temperature": 3.0, "max_weight": 100.0, "target_tau": 0.005, "checkpoint_every": 5000},
"selection": {"seed": 21000, "batches": 2, "candidate_updates": [10000, 20000, 30000], "rule": "lowest final-window mean error among validation candidates satisfying all nominal gates; no selection on test"},
"test": {"seed": 31000, "batches": 4, "boundary_seed": 41000, "boundary_batches": 1, "stability_seed": 51000, "stability_envs": 64, "stability_seconds": 120},
"acceptance": {"mean_error_m": 0.008, "p95_error_m": 0.015, "success_radius_m": 0.015, "success_fraction": 0.95, "baseline_error_reduction": 0.6, "boundary_mean_error_m": 0.012, "boundary_p95_error_m": 0.02, "max_normalized_torque": 1.00001, "max_normalized_velocity": 1.00001, "joint_limit_tolerance_rad": 0.01, "anchor_pose_max_error": 0.0001, "invalid_states": 0, "unexpected_terminations": 0, "memory_growth_fraction": 0.05},
"budget": {"gpu_count": 1, "gpu": "RTX 5080 16 GiB", "wall_seconds": 5400, "disk_gb": 10, "max_training_attempts": 2},
"release_version": "1.0.0"
}
@@ -0,0 +1,79 @@
{
"experiment": "go2_fl_reach_iql_v1_attempt2",
"scope": "Go2 FL only; fixed suspended anchor; 3-DoF foot-link Cartesian stationary reach; nominal reset; no contact, noise, morphology or whole-body transfer claims",
"bundle": "assets/generated/legs/go2_fl",
"max_joints": 8,
"num_envs": 256,
"episode_steps": 200,
"settle_steps": 150,
"goal_action_range": 0.75,
"collection": {
"seed": 12000,
"shards": 20,
"noise_std_by_quartile": [
0.0,
0.03,
0.1,
0.25
],
"feedback_gain": 1.5,
"velocity_gain_s": 0.04,
"integral_gain_s_inv": 8.0
},
"training": {
"seed": 42,
"updates": 30000,
"batch_size": 512,
"hidden": 128,
"lr": 0.0003,
"discount": 0.99,
"expectile": 0.7,
"temperature": 3.0,
"max_weight": 100.0,
"target_tau": 0.005,
"checkpoint_every": 5000
},
"selection": {
"seed": 21000,
"batches": 2,
"candidate_updates": [
10000,
20000,
30000
],
"rule": "lowest final-window mean error among validation candidates satisfying all nominal gates; no selection on test"
},
"test": {
"seed": 31000,
"batches": 4,
"boundary_seed": 41000,
"boundary_batches": 1,
"stability_seed": 51000,
"stability_envs": 64,
"stability_seconds": 120
},
"acceptance": {
"mean_error_m": 0.008,
"p95_error_m": 0.015,
"success_radius_m": 0.015,
"success_fraction": 0.95,
"baseline_error_reduction": 0.6,
"boundary_mean_error_m": 0.012,
"boundary_p95_error_m": 0.02,
"max_normalized_torque": 1.00001,
"max_normalized_velocity": 1.00001,
"joint_limit_tolerance_rad": 0.01,
"anchor_pose_max_error": 0.0001,
"invalid_states": 0,
"unexpected_terminations": 0,
"memory_growth_fraction": 0.05
},
"budget": {
"gpu_count": 1,
"gpu": "RTX 5080 16 GiB",
"wall_seconds": 5400,
"disk_gb": 10,
"max_training_attempts": 2
},
"release_version": "1.0.0"
}
+5 -2
View File
@@ -2,6 +2,8 @@
软硬件描述规范的唯一入口为 [统一软硬件描述规范 v2](software_hardware_spec.md)。本文只记录当前单腿兼容运行链路的实现、验收证据与限制,不是独立规范,也不代表通用 v2 adapter 已实现。
**新增受载站立任务另见[loaded_stand.md](loaded_stand.md)**:可运动底座、六维测力及接地承载,采用新观测契约;本页下述固定支座token和1kg合成anchor描述仅适用于旧链路,不能直接用于移动根。
## 1. 实现范围
```text
@@ -11,7 +13,8 @@ LegSpec JSON → URDF 子树/资源/许可 → USD 与映射
- 已实现真实 Go2 FL、Go1 FL 导入、CUDA 批量仿真、indexed reset、目标条件共享 PPO 短更新。
- 支持独立 bounded revolute 树、显式辅助 fixed 支链、一个足端、sphere/box 间隙探针;mimic、prismatic/continuous、闭链、柔性结构和 mesh 足底近似明确不支持。
- **未实现** ModSkill 式学习 latent、跨部件协调、冻结 SkillBundle、高层调用及真实留出形态泛化。人工 goal、URMA 内部 action intent、保存 checkpoint 都不等于已学得可复用 Skill。
- 第二阶段新增仅绑定 Go2 FL 的离线 IQL Reach 冻结链路,使用本页既有 token/PD 接口;见 [使用说明](stage2_usage.md) 与 [验收报告](stage2_acceptance.md)。
- **未实现** ModSkill 式学习 latent、跨部件协调、通用高层调用及真实留出形态泛化。人工 goal、URMA 内部 action intent、保存 checkpoint 都不等于已学得通用可复用 Skill。
- Go2/Go1 都是 3 DoF;生成的 1/2/4 DoF 夹具验证接口,不代替真实不同 DoF 或迁移实验。
## 2. 资产与物理契约
@@ -64,7 +67,7 @@ LegSpec JSON → URDF 子树/资源/许可 → USD 与映射
不同拓扑按独立同构 worker 采样,只在 PPO 层合并,不能认为 `replicate_physics=True` 支持任意拓扑混装。各资产等量采样、共用采样权重/optimizer,必须同 physics/control dt;device 显式传给 worker。保存未裁剪动作的 log-prob,实际执行再裁剪。更新前检验概率一致性;GAE/旧 rollout 梯度分离,拒绝空/非有限输入。
当前每轮重启 Kit、CPU 文件交换,**仅适合短测试,不适合直接长训或吞吐结论**。checkpoint 仍标 `trained_skill=false`。
本节原 `train_leg_skill.py` 仍每轮重启 Kit、CPU 文件交换,**仅适合短测试,不适合直接长训或吞吐结论**。其 checkpoint 仍标 `trained_skill=false`。新正式路径 `stage2_sim.py` / `train_offline_leg.py` 常驻采集后固定数据离线更新,与旧 PPO 分离,发布有独立证据门槛。
运行报告先写 `passed=false/status=running`,环境成功关闭后才提交成功;JSON 原子替换并拒绝 NaN/Inf。关闭异常/子进程失败/critical log 均保留非零退出码,检查 error/fatal/contact buffer overflow,不能只看进程退出 0。优化按字段读取观测,不缓存跨 reset 状态;FK 常量与 box 投影复用/向量化。
+123
View File
@@ -0,0 +1,123 @@
# Go2 FL:连接可运动底座的承载站立试验
## 状态与范围
已实现独立任务 **`SkillFactory-LoadedLegStand-v0`**,并完成真实PhysX物理验证。预登记范围见[计划](loaded_stand_plan.md),配置为`configs/standing/go2_fl_support_v0.json`。
**本页记录的是环境、六维测力和反馈控制基线,不是新训练/冻结策略的验收。** 用户已将该任务设为当前首要任务,独立PPO首轮训练已完成,原始120秒耐久足端漂移超限;用户随后批准该固定偏差,已带豁免冻结发布,见[训练结果](loaded_stand_training.md)、[原始验收报告](loaded_stand_acceptance.md)及[1.0.0发布报告](loaded_stand_release_1_0_0.md)。 旧Reach 1.0.1、训练源码身份及原固定悬空资产保持不变;新场景没有世界固定根约束,也没有加载旧Reach策略来冒充站立。
## 机械模型
```text
世界参考 ── XY/转动弹簧阻尼导向(z向力严格为0)
│
可运动虚拟底座:3 kg
│
刚性安装关节 + 六维测力
│
传感器载体:0.05 kg
│
Go2 FL
│
足端—地面接触
```
- 根是floating articulation;固定的是**底座与腿之间的安装连接**,不是底座与世界之间的连接。保留所有固定关节,不合并测力边界。
- 导向模拟连接后的侧向/姿态约束:横向刚度250N/m、阻尼20N·s/m;转动100N·m/rad、5N·m·s/rad。导向有限幅,竖直刚度/阻尼/力上限均为0,无重力补偿或竖直悬挂。
- 这是**有导向约束的单腿承重**,不代表完整四足躯干或无辅助单足平衡。`--free-base`可关闭导向做失败探索,未声明该模式稳定。
- 虚拟底座是合成参数,不是厂家躯干模型。原腿真实惯性不改;两条源装饰支链没有惯性,仍采用导入器推定值并披露。运行时实际总质量为**5.388351kg**,不是仅凭已知源质量估算支撑力。
- 平地材料沿用项目地面:静摩擦0.8、动摩擦0.6、恢复系数0;没有验证随机地形/摩擦/传感噪声。
## 六维传感与接口
使用Isaac Lab `JointWrenchSensor`读取PhysX内部固定关节反力,选取child=`anchor_base`的link行,**不是actuator effort,也不是外力指令副本**。
- `mount_wrench`顺序`[Fx,Fy,Fz,Tx,Ty,Tz]`,N/N·m;原点为安装关节,坐标为child joint frame;方向为底座施加在腿上的力/矩。没有重力去皮,包含惯性与接触传递影响。
- 当前安装关节两侧原点重合、child局部旋转为单位旋转。已在实际受力情况下核对符号/力矩原点,不能不经复验推广到任意偏移安装frame。
- 导向力、扰动力都在世界坐标表达,作用于底座安装原点,分别记录;它们不冒充传感值。
- 新契约是`go2-fl-loaded-stand-v0`,不是`leg-token-v1`,也不是通用SoftwareSpec v2 adapter。
`obs["policy"]`各字段为float32、仿真device上的批量tensor:
| 字段 | 形状 | 含义 |
|---|---|---|
| `joint_state` | `[B,3,3]` | q[rad]、dq[rad/s]、上一已执行normalized action |
| `base_state` | `[B,13]` | 相对env原点的位置、**xyzw**四元数(本机Isaac Lab版本)、世界线速度/角速度 |
| `mount_wrench` | `[B,6]` | 安装口实测约束反力/矩,N/N·m |
| `foot_state` | `[B,10]` | 根系足端球心位置、旋转根系相对速度、世界地面法向力3维、gap[m] |
| `guide_wrench` | `[B,6]` | 上一物理子步实际施加的导向wrench,非传感器读数 |
| `goal` | `[B,1]` | 名义站立底座高度[m] |
移动根相对速度包含`omega × r`项;足端接触传感器只提供法向力,不声称摩擦wrench。动作是`[B,3]`,按hip/thigh/calf顺序,裁剪到[-1,1]后转换成`q_nominal + 0.5*action`并裁剪软限位。保留原PD、硬件限矩/限速。奖励包含承重接触、高度、倾角、effort和动作变化;默认有timeout、invalid和fall终止。
## 本轮真实物理证据
最终正试验:`logs/loaded_stand/v0/standing_clean/`;无动力负对照:`logs/loaded_stand/v0/disabled_clean/`。后者的`passed=false`是预期站立失败,没有重标为成功。汇总见`logs/loaded_stand/v0/summary.json`;正试验16项物理/接口检查、进程退出和日志检查均通过,不能只看单个物理报告的passed位。
- **13场景×14秒**:无扰动,加世界±Fx/±Fy/±Fz/±Tx/±Ty/±Tz;幅值5/5/10N、1/1/1N·m。3–6秒与8–11秒施加梯形脉冲,渐入/渐出各0.5秒。
- PGS、500Hz物理、100Hz控制;每个试验7,000批物理快照,13环境合计91,000环境物理步。前2秒settle,试验内无自动复位。
- 初始化前2秒内最大下沉**54.06mm**;下面8.01mm高度结果仅指settle之后,没有隐藏为全程高度误差。若实际装配不允许这种启动下沉,需要另设初始化/控制验收。
- 基线是根据底座位置/速度误差、名义FK Jacobian进行积分修正的关节位置控制器。六维传感器已提供给任务观测并独立验证,但本基线没有使用学习模型或声称学会力反馈策略。
| 指标(13环境最坏值) | 实测 | 预登记门槛 |
|---|---:|---:|
| settle后高度误差 | **8.011mm** | ≤60mm |
| settle后倾角 | **3.292°** | ≤15° |
| 足端中心水平漂移 | **1.562mm** | ≤30mm |
| 承重接触比例 | **100%** | ≥98% |
| 平均竖直力平衡相对误差 | **0.00513%** | ≤10% |
| normalized applied torque峰值 | **0.24634** | ≤1.00001 |
| normalized joint velocity峰值 | **0.18830** | ≤1.00001 |
| 硬限位超出 / 导向z力 | **0rad / 0N** | ≤0.01rad / 0N |
| 安装口准静态力误差 | **0.04027N** | ≤1.5N |
| 安装口准静态力矩误差 | **0.002964N·m** | ≤0.3N·m |
准静态检查独立计算底座重力、COM力矩和实际施加外力,分别比较无扰动[2,3]s、两个载荷平台[4.5,5.5]/[9.5,10.5]s的测量均值。不以这个检查声称动态传感器精度或实机标定。足端中心漂移只是接触位置代理,不是直接摩擦滑移测量。
负对照在**全部13环境**出现约**0.3546m有符号向下位移**,且关节applied torque始终为0,承重接触/高度/力平衡等失败。底座最终可能由自身碰地承载,这不算站立。
`physics_trace.pt`保存每个500Hz子步的底座状态、关节状态/力矩、足端法向力、实际导向/扰动和安装口世界wrench;`trace.pt`另存100Hz控制观测/动作。indexed/empty reset、3控制步timeout的pre-reset terminal observation和复位后传感清零、非法动作原子拒绝均实际验证。reset probe的15个额外物理步不混入7,000步试验统计。
## 失败记录与工程检查
- 首次转换未显式给深层foot开启ContactReportAPI,运行失败;已修复转换器,失败资产保存在`build_initial_asset/`,旧asset没有被改。
- 本资产在原TGS配置下,安装口反力不满足独立静力平衡(约64N误差),不能采信。GPU与CPU调试结果保留`debug_02/03/cpu`。改用本任务独立的PGS配置后通过;不是对整个TGS后端作通用结论,也没有修改旧Reach配置或降低门槛。
- 最终日志审计发现继承的`enable_external_forces_every_iteration=True`是TGS专用,不受PGS支持。已显式关闭并重新执行`standing_clean/disabled_clean`;正试验无critical日志。监督器也新增“即使子进程退出0,只要日志出现Error/Fatal仍拒绝通过”的回归。较早`standing_final/disabled_final`保留,不作为最终干净日志验收。
- 独立只读审查指出倒塌判据应为有符号、逐环境判定,以及应保存500Hz而非仅100Hz历史;两项已修复并重新执行最终正/负物理试验,复审无遗留发现。
- 独立读取500Hz历史重新计算高度/倾角/向下位移、正负载荷峰值和三个测力平衡窗口;发现并纠正了说明/重算器中的四元数顺序假设。本机Lab采用xyzw;仿真实现始终使用同版本Math API,物理结果不变,失败重算记录也保留。
- pytest **207 passed, 3 skipped**,Ruff及`git diff --check`通过。skip为原有可选Kit集成项;本轮真实Kit试验独立执行,不以fixture替代。
- 使用`supervise_sim.py`监督进程组,只清理自己收养的Kit/Hub/telemetry子进程,最终均`cleanup.clean=true`。不声称Kit原生退出没有helper残留。
- 本地新产物约105MiB;单GPU串行执行,未开展训练。最终日志/独立历史重算发现问题后补做复验,开发验证超过原45分钟估计(约50余分钟),不声称满足原时间预算。
## 复现
在仓库根目录,用新输出路径保留所有历史证据:
```bash
unset CONDA_PREFIX VIRTUAL_ENV
export PYTHONPATH="$PWD/source/SkillFactory" PYTHONUNBUFFERED=1
LAB=/home/cen/IsaacLab/isaaclab.sh
# 仅在新机器/不存在派生资产时执行;拒绝覆盖已有目录
python3 scripts/supervise_sim.py --log logs/loaded_repro/build.log \
--report logs/loaded_repro/build_process.json -- \
"$LAB" -p scripts/assets/build_loaded_leg.py --headless
# 正试验
python3 scripts/supervise_sim.py --log logs/loaded_repro/stand.log \
--report logs/loaded_repro/stand_process.json -- \
"$LAB" -p scripts/verify_loaded_leg.py --num_envs 13 --duration 14 \
--output logs/loaded_repro/stand --headless
# 负对照:预期非零退出;必须检查negative_control_collapsed和逐环境向下位移,
# 不能把任何异常退出都当作“符合预期的倒塌”。
python3 scripts/supervise_sim.py --log logs/loaded_repro/disabled.log \
--report logs/loaded_repro/disabled_process.json -- \
"$LAB" -p scripts/verify_loaded_leg.py --num_envs 13 --duration 14 \
--controller disabled --output logs/loaded_repro/disabled --headless
```
去掉`--headless`可观察GUI。少于13环境或短于14秒只能调试,完整协议门槛会失败;单环境演示可加`--scenario-offset 1`观察横向扰动。修改载荷质量/几何时需更换配置中的`support_asset`到新目录并重新生成,不能让旧USD与新质量配置混用。
下一阶段需要单独定义正式RL训练、载荷/刚度分布、留出扰动和新冻结包验收。当前结果不覆盖大冲击、随机底座运动、其他腿、整机平衡、承重行走或实机安全。
+52
View File
@@ -0,0 +1,52 @@
# Go2 FL受载站立:1024候选独立验收预登记
**后续状态:用户已明确批准本次固定漂移偏差,现已带豁免冻结发布1.0.0,见[发布报告](loaded_stand_release_1_0_0.md)。以下原始门槛、实测失败和首次拒绝结论仍按发生时的事实保留,不回填成功。**
用户授权独立留出测试、完整物理验收和冻结发布。仅测试已选择的1024更新,不重训、不使用holdout重新选模或改变门槛。候选checkpoint SHA256为`f2197339dea3622ed92b82fdddff3400f2a82f3ca4d4fd3be6c72b45f1a77704`。
## 固定协议(首次留出运行前登记)
- 新发布名预留`go2_fl_loaded_stand-1.0.0`,不是Reach后继版本。旧训练源码、配置、资产、权重及旧发布保持不变。先提取纯actor、固定normalization、绑定task/support hash生成candidate manifest;所有物理证据绑定整个candidate身份。actor纯PyTorch直接3动作,无teacher、残差、critic或采样噪声。
- **holdout**:seed83000,256环境×14秒;13场景平衡覆盖后独立打乱,幅值[0.6,1.0]、时移[-0.25,0.25]。同一输入的独立零动作hold对照;settle2秒后逐500Hz高度均值比hold改善≥30%。所有环境仍须满足原高度/倾角/foot drift/contact/vertical balance及全时段安全阈值。禁用自动reset、完整记录失败状态;出现fall不会被重置后的好状态替代。
- **full**:13环境×14秒,原默认13类单轴±全幅、零时移、seed71000;必须通过既有16项环境物理门槛,额外记录高度均值/每环境结果及全时段跌倒。原测力三个准静态窗口[2,3)、[4.5,5.5)、[9.5,10.5)秒,force≤1.5N、torque≤0.3Nm,不进行偏置拟合。
- **stability**:26环境×120秒(每场景2环境),seed85000;每14秒重复完整载荷协议,不reset,2秒只在开头settle一次;整个后续118秒沿用高度≤0.06m、倾角≤15°、接触≥98%、足端水平漂移≤0.03m、平均竖直平衡误差≤10%。所有物理子步有限、torque/velocity≤1.00001、硬限位超出≤0.01rad、guide Fz严格0,无跌倒/异常reset。复用每周期准静态窗口校验测力。GPU进程显存尾部10秒峰值相比2–12秒窗口峰值增长≤5%,外部nvidia-smi采样并计入真实物理时间;不以空尾部报告0增长。
- **exploratory**:60环境×14秒,seed84000;所有两轴组合15对×4符号、各轴全幅,单独披露,非训练分布、非发布适用域。不可用于选模。探索失败保留,不将其重标通过,也不把单轴域的通过泛化为组合域能力。
- 全部模式保留500Hz原始物理trace和稀疏实际观测/动作;CPU独立重算所有物理门槛和原始载荷覆盖;不能仅信报告passed。hold基线可能不通过技能门槛,必须如实保存退出非零和失败门槛,不阻止以相同输入评估其均值对照。
- 适用物理域不变:PGS、floating root、3kg合成底座、有效总质量约5.388351kg、XY/转动导向,竖直辅助0;不是自由单足平衡、四足、跨形态或实机安全认证。foot drift仅球心滑移代理、contact仅法向力。源缺失惯性/导入推定值继续披露。
## 工程与冻结
- 新接口`go2-fl-loaded-stand-v0`,45维固定SI输入,xyzw四元数;与旧leg-token-v1及通用v2 adapter不兼容。
- 对candidate及最终冻结目录/归档解包执行真正隔离CPU解释器package-only加载(复用已有Python3.12/Torch2.10 CPU venv,不安装运行时),B1/16/64/256各100次确定性调用、输入/权重/buffer不变、无grad/训练/仿真依赖;错误shape/dtype/device/finite/绑定和损坏文件拒绝。
- 实际物理观测的CPU/CUDA输出最大绝对差≤1e-5。检查点actor提取与训练策略确定性输出等价、critic/optimizer/RNG不进推理包。
- 发布门禁fail-closed:完整训练/validation-only选模、候选身份、独立加载、holdout和对照、full、stability及原始trace重算、资源/工程检查、干净进程日志全部通过才原子发布;没有足够证据不得写accepted。探索结果必须存在并如实披露,但不是单轴域门禁。
- 最终版本目录独占、禁止覆盖;最终manifest新增证据时明确绑定已测试candidate manifest与不变运行时核心,不偷换测试身份;暂存最终包通过独立加载后才原子rename,再验证正式目录和归档。SHA256是完整性校验,不是数字签名。
- 资源护栏:本轮实际仿真/加载累计上限30分钟、额外产物2GiB(不含开发时间);不自动开展第二候选或调参搜索。发现工程错误保留失败证据,仅对同一候选修复验证工具并记录新执行身份;任务/验收门槛不变。
证据根目录:`logs/loaded_stand/acceptance_v1/`。以上是首次留出运行前的预登记,门槛未变。
## 执行结果:耐久门槛失败,拒绝发布
最终证据在`logs/loaded_stand/acceptance_v1/v2/`。独立审查修正了每周期测力窗口必须分别取误差的问题,并补全包内混合坐标系说明;原candidate及第一批测试完整保留。说明修正改变manifest,因此同一权重、同一输入协议重新执行full/holdout/hold;这是同一留出集的身份复验,不是新的独立测试集,也没有以测试结果选模。
固定checkpoint SHA未变;actor SHA256始终为`1384b5aa780cded1f4ed21b617ff8c883a5f82826c4cb4c9415b75f315b1e664`。修正说明后的candidate manifest为`7cb478a3b17aaf307d2d6e3c2b49cc25aa9af1ee9d118dd731b52f618c27e628`,不是已发布版本。
| 测试 | 结果 | settle后平均/最大高度误差 | 最大足端漂移 |
|---|---|---|---|
| 独立holdout,256×14秒 | 通过 | 0.921 / 10.693mm | 4.442mm |
| 同输入零动作hold | 高度不通过,保留失败 | 66.377 / 97.808mm | 3.313mm |
| 原13类全幅,13×14秒 | 全部19门槛通过 | 1.097 / 11.397mm | 4.637mm |
| 连续耐久,26×120秒 | **foot_drift失败** | 1.703 / 20.279mm | **30.783mm >30mm** |
| 两轴组合探索,60×14秒 | 本次探索通过,不泛化适用域 | 1.673 / 19.595mm | 7.915mm |
- holdout相对hold平均高度误差改善**98.6126%**。各试验最低承重接触比例均100%、无跌倒/自动reset;guide Fz严格0、物理/测力有限。
- 耐久环境6/19都是**负向Fz 10N**场景;漂移分别30.620924/30.782583mm,首次越界在115.828/115.760秒,峰值在119.054/119.044秒。参考点固定在2秒,不按载荷周期重置。
- 父级独立CPU读取全部500Hz trace并重算;另用逐点Python `math.hypot`独立重算两个失败环境,峰值与报告差<1e-12m。不是报告舍入或末段漏采,不能以“仅超0.78mm”为由放行。
- 耐久其余门槛通过:最大倾角0.057942rad,torque/velocity比例0.265703/0.555917,竖直平均平衡误差≤1.06e-6;26个独立准静态窗口各500样本,最大测力残差0.029747N/0.001857Nm;2–12秒与110–120秒显存峰值增长0%。
- candidate在真正隔离CPU解释器完成B1/16/64/256各100次确定性调用/坏输入和损坏拒绝;真实物理观测CPU/CUDA最大动作差2.3842e-7。没有声称最终包/归档已经独立验证,因为没有生成正式包。
- 五个最终试验全部完成和清理;hold与stability保持原失败/非零退出,唯一critical记录是launcher对预期失败的退出报告,没有后端Error。探索是在必需耐久失败后单独完成的已登记信息性试验,不用于放行。
- 全部物理尝试(含旧说明身份复验)累计**502.72秒**;原始证据约**1.72GB**,符合30分钟/2GiB护栏。旧Reach发布全部checksum通过,原训练源码与候选权重不变。
`v2/trace_audit.json`的passed仅表示证据读回一致,其`accepted_skill=false/publication_blocked=true`;`v2/publication_gate.json`实际非零拒绝必需试验失败。未调用正式发布操作、未创建`releases/go2_fl_loaded_stand-1.0.0/`。发布器另补了暂存/最终加载/归档加载/提交后报告失败的隔离回滚,以及锁内不可覆盖检查;这些纯工程测试不代替真实成功发布。
**下一步需用户授权新一轮训练**,针对负向竖直载荷下长期足端漂移改进训练/长回合覆盖。不得用已查看的holdout重选本轮其它checkpoint、缩短耐久时长、重置漂移参考点或放宽30mm门槛。新一轮应独立预登记,并保留本轮失败。
+32
View File
@@ -0,0 +1,32 @@
# 单腿连接底座承载站立:v0实施/验收计划
## 范围(实现前登记)
Go2 FL单腿,地面接触,可运动的3kg虚拟底座→保留的固定测力关节→0.05kg合成传感器载体→原腿。世界根固定约束必须不存在,`is_fixed_base=false`。真实腿源惯性不改;源中两条装饰支链惯性缺失,保留导入器推定值并报告实际质量,不冒充厂家惯性。
默认采用弹簧阻尼导向,横向250N/m、20N·s/m,转动100N·m/rad、5N·m·s/rad;世界z向刚度/阻尼/外加导向力均为0。它模拟其他装配结构的侧向/姿态约束,不是完整四足躯干,也不是无辅助单足平衡。脚必须承担竖直重量;无动力负对照应倒塌。可将导向关闭做自由底座实验,但不预先声称稳定。
挂载传感器取PhysX JointWrenchSensor的内部固定关节反力,单位N/N·m、顺序Fx,Fy,Fz,Tx,Ty,Tz,原点为mount joint、坐标为child joint frame。不是将施加的扰动力复制成传感读数,不做隐式重力去皮。独立静力平衡确认符号和力矩原点。世界坐标下的导向wrench与扰动指令独立记录。
新任务/观测契约为`go2-fl-loaded-stand-v0`,不兼容Reach冻结权重。先交付物理环境、传感器、奖励/失败条件、站立基线与实测报告;**本轮不是新的RL训练/冻结发布**。旧任务、旧资产、已发布1.0.0/1.0.1及训练绑定源码不改。
## 预定试验与门槛
配置见`configs/standing/go2_fl_support_v0.json`。500Hz物理/100Hz控制,14秒episode,前2秒settle。13类场景为无扰动和世界坐标±Fx/±Fy/±Fz/±Tx/±Ty/±Tz,幅值5/5/10N、1/1/1N·m,作用点为虚拟底座安装原点。3–6秒及8–11秒施加梯形脉冲,0.5秒渐入/渐出。完整检查至少13环境,允许少量环境作调试但不算全类别验证。
- 检查完全浮动root、固定sensor joint保留、足端地面接触,传感器不是零值占位。
- 每物理子步:有限状态、硬关节越界≤0.01rad,归一化torque/velocity≤1.00001;保存实际导向、扰动和六轴测量。
- settle后每环境:高度误差≤0.06m、倾角≤15°、足端中心水平漂移≤0.03m、有效承重接触比例≥98%。足端中心漂移是接触位置代理指标,不等于摩擦滑移测量。
- 竖直合力统计:平均足端法向力与实际总质量*g减实际施加竖直外力之差≤10%;导向z力严格为0。传感器静力力误差≤1.5N、力矩误差≤0.3N·m(使用同原点/同坐标)。
- 负对照:关闭关节驱动力,底座应明显下落/跌倒;不能靠底座弹簧把腿悬挂起来。
- 只读审计→纯数学/配置测试→小规模真实物理调试→完整13类试验及负对照→回归。失败和调参过程保留,不将调试当holdout或正式学习验收。
- 单GPU;本轮最多45分钟开发/物理验证资源窗口,不开展大规模训练。到预算则如实报告未完成项。
需要正式训练时另定数据分布、质量/刚度变化范围、未知扰动评估和冻结接口;不能仅以奖励上涨或本轮基线通过宣称训练完成。
## 实施补记(不改上述门槛)
- 实际传感读数在该资产的TGS配置下未通过独立平衡检查,改用本任务独立PGS配置;失败日志保留,不以修改符号或放宽误差阈值掩盖问题。
- 准静态传感检查增加至3个窗口:无载荷平台[2,3]s和受载平台[4.5,5.5]/[9.5,10.5]s,仍使用原力/矩容差。
- 按独立审查补全500Hz历史,并将负对照判据明确为全部环境向下位移≥0.10m且电机力矩为0;增加真实timeout终端状态/复位与非法动作拒绝检查。
- 实施结果与边界见[实测报告](loaded_stand.md);这不是新RL策略发布。
+51
View File
@@ -0,0 +1,51 @@
# Go2 FL受载站立1.0.0:带明确偏差豁免的冻结发布
## 当前授权与范围
用户在获知120秒耐久漂移30.62/30.78mm超过30mm后明确表示:
> 这个120秒耐久的飘逸超过的很小,可以视为通过
据此只对**当前1024候选、当前120秒耐久原始证据、环境6和19的foot_drift一项**作特例放行。不是将全局阈值改成31mm,不授权其它失败、其它权重或新的试验结果,不代表实机安全认证。
- 原始结果:环境6=0.030620924223339147m;环境19=0.03078258298377699m;原限值0.03m。
- 原始`stability/result.json`及`publication_gate.json`仍为失败,旧summary保留。
- 候选manifest:`7cb478a3b17aaf307d2d6e3c2b49cc25aa9af1ee9d118dd731b52f618c27e628`。
- checkpoint:`f2197339dea3622ed92b82fdddff3400f2a82f3ca4d4fd3be6c72b45f1a77704`。
- 原始holdout/full/exploratory/耐久trace均不重跑、不改写;继续独立CPU重算与发布完整性检查。
- 豁免记录:`logs/loaded_stand/release_deviation_v1/operator_waiver.json`,SHA256 `3c8c03efcedc3c6eb2a0a8ca8fdae2629834473fb122b2ce5bbcc39161344d1e`。
本次使用`--waiver`与`--waiver-sha256`双显式参数。缺少批准、hash不符、额外失败、实测值/证据变化、不同候选均拒绝;不传豁免时原发布拒绝行为不变。软件只返回单独的豁免回执,不能将原始passed或门槛改成true。
正式包必须明确标记`accepted_with_deviation`、`acceptance_basis=operator_waiver`及`unwaived_required_physics_passed=false`,保留实际`foot_drift=false`;附带`DEVIATIONS.md`和原文豁免记录。其它物理/源码/日志/加载门槛保持不变。最终目录及归档必须在独立CPU解释器中验证。
## 发布状态
**已完成带偏差豁免的冻结发布,不是原始物理门槛全部通过。**
- 只读目录:`releases/go2_fl_loaded_stand-1.0.0/`,30个manifest条目;177,670字节归档:`releases/go2_fl_loaded_stand-1.0.0.tar.gz`。
- 最终manifest SHA256:`27e5e56908a0a8585a13913f27663b0dd3cdb7270614a0a05355334de4093e98`。
- 归档SHA256:`ad59972e041f079cc2584923aaa15798857e5619488ba76b69835b558956bd76`。
- 验收报告:`releases/go2_fl_loaded_stand-1.0.0.acceptance.json`;同时保留`staging-load`、`independent-load`、`archive-load`报告。候选、暂存、正式目录和解包均在隔离CPU解释器中通过独立加载;B1/16/64/256各100次确定性调用、损坏/接口错误拒绝。
- 测试:256 passed、3个原可选Kit测试skipped;Ruff/diff通过,独立代码复审OK。增加无豁免仍拒绝、错误范围/hash/额外失败拒绝,以及审批读取/复制时替换再恢复的竞态回归。
- 未重训或重跑物理,不改actor/normalization/接口/候选manifest。原始拒绝报告、原始trace、旧summary及旧Reach发布校验均不变;新增全局checksum条目而非覆盖旧条目。
- `logs/loaded_stand/release_deviation_v1/final_verification.json`再次核对只读权限、核心字节、批准记录、归档全部manifest文件和旧证据;`raw_stability_passed=false`被明确保留。
## 使用
将归档解包,在Python3.12/Torch2.10.0环境中导入包内`frozen_standing.py`;不需要Isaac、项目代码、训练checkpoint或NumPy。部署方必须从实际rig配置获得task/support hash,不能用包内期望值替代对部署rig的核对。
```python
from frozen_standing import FrozenStanding
policy = FrozenStanding(
"go2_fl_loaded_stand-1.0.0",
manifest_sha256="27e5e56908a0a8585a13913f27663b0dd3cdb7270614a0a05355334de4093e98",
task_sha256=deployed_task_sha256,
support_sha256=deployed_support_manifest_sha256,
device="cpu",
)
action = policy.act(observation) # float32[B,3],100Hz
q_target = policy.position_targets(observation) # rad,驱动仍需限矩/限速
```
完整混合坐标系/单位见包内README;已批准的漂移偏差见`DEVIATIONS.md`。适用域仍是Go2 FL、3kg合成底座、PGS、有XY/转动导向且竖直辅助为0的仿真单腿,不是无辅助平衡、四足或实机安全认证。
+73
View File
@@ -0,0 +1,73 @@
# 当前首要任务:Go2 FL受载站立策略训练
用户已要求将受载站立设为主线并开始训练。旧Reach1.0.1仅维护归档/完整性,不再作为当前开发重点。任务物理契约及既有基线见[loaded_stand.md](loaded_stand.md)。
## 首轮结果:训练完成,随后耐久验收失败
2026-10-12首轮完整结束:**2,048次更新、16,777,216条在线转移**,含启动/清理共**1,510.50秒(25.18分钟)**;单GPU、无恢复重启,run约12.6MB,满足本轮90分钟/2GiB预算。KL早停按预登记执行,实际14,979次optimizer minibatch、30,676,992条梯度抽样,不把未执行的epoch计入训练量。
固定validation(256环境×14秒,seed82000)结果:
| 指标 | 选中1024更新 | 零动作hold基线 |
|---|---:|---:|
| settle后平均高度误差 | **0.9663mm** | 66.4361mm |
| 最大selection高度误差 | 11.4227mm | 99.4814mm |
| 最大倾角 | 0.039608rad | 0.052887rad |
| 跌倒比例 | 0 | 0 |
| 最低环境承重接触比例 | 100% | 100% |
按validation-only规则选中**1024更新**,平均高度误差改善**98.55%**;最终2048更新误差2.5967mm,并未默认选最终权重。8次validation与hold使用完全一致的输入;hold是零动作控制对照,**不是holdout**。现有控制基线8.01mm是另一个协议下的最大误差,不能与此处均值直接对比。
- 候选:`logs/loaded_stand/train_v1/checkpoints/best_validation.pt`。
- 编号副本:`checkpoints/update_001024.pt`;已读取并核实model和optimizer逐tensor一致。两文件因保存时机/RNG状态等元数据不同,不要求整个文件hash相同。
- 候选SHA256:`f2197339dea3622ed92b82fdddff3400f2a82f3ca4d4fd3be6c72b45f1a77704`;后续验收须绑定此候选,不以holdout重新选模。
- 训练期间共3次跌倒终止;所有8次validation未跌倒。全程物理/传感有限,关节硬限位超出0、导向Fz为0,所有12种有符号载荷均实际出现。normalized torque峰值0.666229、velocity峰值1.000000477,在1.00001容差内,**不声称速度有裕量**。
- 最终累计390,680批物理快照,其中训练327,680批;其余是hold/validation,不混入在线训练转移量。进程退出0、critical日志为空、自有helper清理完成。
- 父级完成检查读取了4份checkpoint、检查185个tensor,确认真实参数变化、固定归一化buffer不变,并做了CPU合成输入确定性调用。**这只是检查点完整性/数值检查,不是包-only加载或物理验收。**
- 原绑定源码与旧Reach发布校验均未变。完整证据:`result.json`、`selection.json`、`validation/`、`completion_verification.json`、`completion_monitor.md`,以及外部`train_v1_process.json`。
**后续独立验收已执行**:holdout与原13类全幅通过,候选包独立CPU加载通过;但26环境×120秒耐久的两个负Fz场景足端漂移30.62/30.78mm超过30mm门槛,**拒绝冻结发布**。详见[验收报告](loaded_stand_acceptance.md)。该次原始验收的`accepted_skill=false`保持不变,不重选权重、不自动重训。**之后用户明确批准这一次漂移偏差,已另行带豁免冻结发布1.0.0**,见[发布报告](loaded_stand_release_1_0_0.md);不能把该决定回写成原始物理门槛通过。
## 首轮预登记(在正式梯度训练前固定)
- 新策略从随机初始化开始,**在线PPO**,直接输出3个关节位置目标的归一化动作;不加载Reach权重、不把解析站立控制器作为策略输出或隐藏残差。
- 输入新任务全部45维观测,包括真实安装口六维wrench;固定物理量缩放、base z减目标高度,无训练中滚动统计。两层128维actor/critic,tanh-squashed Gaussian;PPO概率比使用pre-tanh latent(同一变换的Jacobian抵消)。不声称URMA共享形态或通用v2 adapter。
- 256环境,32控制步rollout,2048次PPO更新:**16,777,216条在线转移**。每更新4 epochs、minibatch2048;KL超过0.02时停止该更新后续minibatch,不改变转移预算。其余数值锁定`configs/training/go2_fl_loaded_stand_ppo_v1.json`。
- 沿用已校准的PGS、3kg载荷、XY/转动导向及零竖直辅助。每次episode随机选择13种有符号轴向场景,幅值为配置上限的[0.5,1.0]、脉冲时移[-0.5,0.5]秒;不增加未验证的质量/刚度/摩擦随机化。
- 环境奖励除以control dt恢复逐步尺度;真正终止额外-5。GAE对timeout使用**复位前terminal observation** bootstrap,对真正跌倒置零,episode边界切断递归;不得用复位后状态偷接回报。
- 每128更新保存完整model/optimizer/CPU+CUDA RNG/更新计数/累计时间/配置与资产及源码hash;保存原子替换。中断恢复保留learner状态但重置物理episode,**不声称PhysX逐位恢复**。`STOP`文件请求在rollout/update边界保存并退出。
- 每256更新进行独立确定性validation(seed82000,全部256环境,14秒;幅值[0.6,1.0]、时移[-0.25,0.25])。以`平均高度误差 + 0.10m×跌倒比例`最小选模,失败后的剩余步按0.10m误差计,避免早结束获得虚假低误差。相同validation输入重复比较,validation不回传梯度;训练RNG在验证前后恢复。
- 启动前记录固定零动作hold基线;validation仅负责选模,**不是发布验收**。最终选择锁定后才允许读取seed83000 holdout:独立扰动采样并加入两轴组合探索(超出本轮训练分布,单独报告),不得以holdout重选权重或调整门槛。
- 后续候选必须在默认13类全幅协议下逐物理子步通过既有站立/安全/测力门槛,并完成新接口独立加载/耐久等冻结验收;至少比hold基线的平均高度误差改善30%。训练完成、奖励提高或validation最优都不等于技能已发布。
## 本轮预算及失败处理
单GPU,正式首轮最多90分钟wall time、最多2GiB新run产物;到预算保存并标记paused,不将未完成2048更新标成完整训练。监督进程额外留5分钟启动/清理宽限。保留所有预检和失败尝试,不启动无限调参/自动多轮训练。纯数学/恢复单测与短Kit预检只作工程准备,不计入正式训练量。
日志和权重位于`logs/loaded_stand/train_v1/`,预检在独立`logs/loaded_stand/train_preflight/`。`progress.json`是当前进度,`metrics.jsonl`记录逐更新训练/validation;checkpoint不是可直接发布的技能包。训练后还需独立物理测试,不改变既有冻结包。
## 启动记录与操作
2026-10-12首轮已启动并完成,当前没有继续运行的训练进程。启动门禁证据在`train_preflight/`:225 passed/3 skipped、Ruff通过;`run_02`在第2更新真实退出,再通过新Kit进程恢复至第4更新并完成validation。4个预检更新各发生156次正常timeout,覆盖terminal bootstrap和reset;固定下一批数据的learner/optimizer/RNG精确恢复另有纯Torch回归。预检只有工程意义,不计入正式1678万转移。独立审查的验证评分问题已修复、测试并复审通过。
```bash
# 正式首轮已完成;原目录保留,不要覆盖或在原run上追加训练。
# 新的独立复现应替换下面所有train_v1输出名。
unset CONDA_PREFIX VIRTUAL_ENV
export PYTHONPATH="$PWD/source/SkillFactory" PYTHONUNBUFFERED=1
export CUBLAS_WORKSPACE_CONFIG=:4096:8
python3 scripts/supervise_sim.py \
--log logs/loaded_stand/train_v1.log \
--report logs/loaded_stand/train_v1_process.json --timeout 5700 -- \
/home/cen/IsaacLab/isaaclab.sh -p scripts/train_loaded_stand.py \
--config configs/training/go2_fl_loaded_stand_ppo_v1.json \
--output logs/loaded_stand/train_v1 \
--console-log logs/loaded_stand/train_v1.log --headless
# 请求安全暂停;不会直接杀死其他Kit/桌面进程
# touch logs/loaded_stand/train_v1/STOP
```
`train_v1.launch.json`保存启动命令/监督PID。监督报告仅在退出及清理后最终落盘;运行中不能把`passed=false/status=running`当作失败。结束后同时检查训练`result.json`、`train_v1_process.json`及日志,不能以进度文件或子进程退出0代替完整判断。
恢复时先确认原进程已退出、保存并移走STOP文件;读取`latest_checkpoint.json`,使用同一config/output和`--resume <实际checkpoint路径>`,为新监督log/report选择未占用文件名。不要改绑定源码/配置后强行恢复;从checkpoint重启会重置物理episode。程序没有自动重启/自动换参数/自动发布。
+2 -2
View File
@@ -6,7 +6,7 @@
**已落地:** PartSpec v2 / SoftwareSpec v2、JSON Schema、跨文件语义校验、不同部位的设计示例、现有 Go2 单腿描述迁移与回归测试。仓库只保留本规范及两份 v2 Schema,不再并列维护旧版规范。
**尚未落地:** v2 通用仿真/训练 adapter、描述驱动 launcher、实机驱动、冻结 SkillBundle、跨部件协调训练。设计模板不是厂家资产;静态契约匹配不证明机械结构可制造、物理模型准确或策略已训练。当前可运行的仍是 [既有单腿任务](leg_interface.md);该文档仅记录实现和运行验收,不是另一份软硬件描述规范。
**尚未落地:** v2 通用仿真/训练 adapter、描述驱动 launcher、实机驱动、v2 通用冻结发布、跨部件协调训练。第二阶段已有绑定旧 leg-token-v1 的独立 Go2 Reach SkillBundle 流程,见 [验收报告](stage2_acceptance.md),不改变本规范 v2 的运行支持边界。设计模板不是厂家资产;静态契约匹配不证明机械结构可制造、物理模型准确或策略已训练。当前可运行的是[既有单腿任务](leg_interface.md)及专用的[受载站立实验任务](loaded_stand.md)。后者使用独立`go2-fl-loaded-stand-v0`观测契约和六维安装口传感器,仍不是v2通用adapter;已完成独立PPO首轮训练和物理测试,并在用户明确批准本次耐久漂移偏差后[带豁免冻结发布1.0.0](loaded_stand_release_1_0_0.md);原始30mm漂移门槛并非通过,不代表v2通用能力或实机安全。这些文档仅记录具体实现和运行验收,不是另一份统一软硬件描述规范。
## 2. 分层模型
@@ -89,7 +89,7 @@ SoftwareSpec v2(要求、容量、语义端口、控制模式、adapter、产
### 5.1 身份、绑定和能力要求
`kind=part_software`,身份用 `software_id/software_version`;当前交付阶段为 design/prototype,`trained_skill=false/frozen=false`。本版本不提供已训练/冻结发布流程,未来需要单独制定证据门槛。
`kind=part_software`,身份用 `software_id/software_version`;当前 v2 格式交付阶段为 design/prototype,`trained_skill=false/frozen=false`。本版本不提供已训练/冻结发布流程。第二阶段的独立 `skillfactory.skillbundle.v1` 有自己的证据门槛,不伪装成已获运行支持的 v2 SoftwareSpec。
`hardware_bindings` 是可替换候选列表,`binding_mode=alternatives`:每个部件**独立满足**同一 profile,不能把候选 A 的传感器与候选 B 的执行器合并满足要求。不是同时装配,也不代表已经学会跨部件迁移。
+83
View File
@@ -0,0 +1,83 @@
# 第二阶段历史验收报告 — Go2 FL Reach 1.0.0
**最新交付为[1.0.1完整边界复验与发布报告](stage2_release_1_0_1.md):12项检查全部通过,已实际发布。** 下文只记录1.0.0,不将新结果追写入旧版证据。
**对1.0.0的复核结论:训练、恢复、选模和冻结推理闭环保留有效;历史报告不足以证明修正后的完整发布验收。** 后续发现并修复了工程/恢复门槛遗漏、边界覆盖不全、失败复验残留成功报告三项问题,见 [修复记录](stage2_reliability_fix.md)。旧边界实际仅覆盖2/6个面、4/12条边、8/8个角。
以下是2026-10-12的**历史验收记录**。原机器报告`logs/stage2/v1_attempt2/acceptance.json`及1.0.0包内`accepted`保持原样以便追溯,**不代表已通过修正后的门槛**。旧版“全部完成”结论据上述复核限定。不是通用部件策略或整机迁移完成声明。
## 1. 交付
- 冻结包:`releases/go2_fl_reach-1.0.0/`;归档:`releases/go2_fl_reach-1.0.0.tar.gz`;完整性入口:`releases/SHA256SUMS`。
- 正式配置:`configs/training/go2_fl_offline_v1_attempt2.json`。
- 代码:`legs/offline.py`、`legs/frozen_policy.py`;`scripts/{stage2_sim,train_offline_leg,export_leg_bundle,verify_frozen_leg,verify_stage2_delivery}.py`。
- 测试:`tests/test_offline_leg.py`;使用/复现:[stage2_usage.md](stage2_usage.md);预注册:[stage2_plan.md](stage2_plan.md)。
- 第二轮全量数据、训练恢复点、日志和评估:`logs/stage2/v1_attempt2/`。第一轮失败产物完整保留在 `logs/stage2/v1/`,没有覆盖或冒充通过。
## 2. 任务与方法
真实 Go2 FL 三关节、固定悬空支座,500 Hz PhysX / 100 Hz position-PD 控制。目标为 anchor-frame **足端 link** xyz/L,由 `q_nominal + 0.25*a` 的 FK 生成,`a∈[-0.75,0.75]^3`。每个独立目标从名义姿态复位,2秒运行,最后0.5秒判定。成功要求该窗口**每步**误差≤15mm;均值/P95以episode末段平均误差为统计单位。
IQL:固定数据、双Q TD、expectile V、优势加权确定性actor。策略只接收现有token投影的15个状态/目标数;不接收目标关节角。采集器可使用目标关节角、PD反馈、积分补偿与噪声,**发布策略不包含采集器、IK或积分控制器**。仅actor及固定训练集mean/std用于推理。
## 3. 正式训练、恢复与选模
- 20 shards ×256环境×200步 = **1,024,000条**固定转移,5,120个训练目标;训练种子42,采集种子12000–12019。
- 完成 **30,000次**更新,batch512(15,360,000次转移抽样),没有以短更新/reward上升作为验收。
- 第5,000步真实退出并从完整模型/target/optimizer/CPU与CUDA RNG恢复,继续到30,000步。两个独立进程从同一真实5k检查点恢复至5001,全部learner/optimizer/RNG逐tensor一致;另有连续更新vs保存恢复的单元回归。
- 独立validation种子21000–21001,512目标;10k/20k/30k平均误差分别 **2.621/2.211/2.222mm**,全部过门槛。按预定最小平均误差规则选择 **20,000步**,不是默认选最终权重。
- 首轮完整30k训练因行为数据稳态误差过大而未通过(10k候选均值10.45mm、成功83.0%)。第二轮修改记录见 [stage2_attempt2.md](stage2_attempt2.md)。阈值未降低,test种子未用于调参。
## 4. 独立效果验收
| 项目 | 实测 | 预注册门槛 |
|---|---:|---:|
| 未见目标,seed31000–31003 | 1024 episodes | 1024 |
| 末段平均误差 | **2.184mm** | ≤8mm |
| episode均值P95 | **4.532mm** | ≤15mm |
| 成功率(窗口最大误差≤15mm) | **100%**,Wilson95% CI [99.626%,100%] | ≥95% |
| 零动作同目标基线均值 | 53.019mm | 对照,非学习目标 |
| 相对基线误差下降 | **95.881%** | ≥60% |
| 旧边界目标,seed41000(面/边覆盖不全) | 256 episodes | 256 |
| 边界均值 / P95 | **2.709 / 4.719mm** | ≤12 / ≤20mm |
| 边界成功率 | 100%(补充指标) | 不以此替换边界均值/P95 |
测试和边界报告逐episode保留均值及最大误差。验证/测试是独立进程,不复用训练环境状态。包审查发现需要绑定**整个manifest**而不只是权重,因此相同已选权重/目标/门槛又完整重跑一次,仅补全证据身份;此前结果保留 `pre_manifest_evaluation/`,没有在看过测试后调参或重新选模。
## 5. 运行稳定性与控制约束
最终耐久:**64环境×120秒**、12,000控制步/768,000环境控制步,另200步warmup;每个物理子步取样,共61,000批物理快照。每2秒换目标,不复位,验证持续运行安全而非连续轨迹跟踪精度。
- 非有限状态、意外termination/reset:**0**。
- 全套最终测试/边界/耐久:关节硬限位超出量 **0 rad**,支座pose漂移 **0**。
- 耐久峰值归一化applied torque **0.663709**(阈值1.00001)。
- 耐久峰值归一化速度 **1.000000834**(阈值1.00001):触及速度上限,只有数值容差范围内超量,**不宣称存在速度安全裕量**。
- warmup后总设备显存(含PhysX/Warp)初始/峰值均 **3,544,711,168 bytes**;增长 **0%**(阈值≤5%),包括最后一次采样。
- 最终collect、validation、test、boundary、stability日志无 `[Error]/[Fatal]/buffer overflow`;环境关闭成功才提交passed。
这里的torque是仿真控制器applied torque,不是实测电机力矩;无实机安全、watchdog、承重、足底摩擦wrench或接触控制能力声明。
## 6. 冻结与独立加载
独立版本域:`skillfactory.skillbundle.v1` / `go2-fl-reach-leg-token-v1` / 1.0.0;**不冒充 SoftwareSpec v2通用adapter**。包内含actor、归一化buffer、完整token schema、静态token/物理绑定、关节顺序/PD/限位、版本依赖、示例输入、来源许可、选模/评估证据和每文件SHA256。
- 必需推理依赖:Python3.12 + Torch2.10.0及pip传递依赖;不需要NumPy、Isaac/Kit、项目导入、资产路径、数据集或训练checkpoint。
- 全新CPU venv(Python3.12.3/Torch2.10.0+cpu),将包复制到独立临时工作目录,清除PYTHONPATH/Isaac变量,`python -I`启动;B=1/16/64/256,共400次重复调用通过。
- eval + requires_grad=false + inference_mode;参数/所有buffer和调用者输入调用前后逐tensor不变。无optimizer,无统计更新路径,无内部递归状态。
- 错hardware hash/part、字段/shape/dtype、有效槽NaN/Inf、mask/context、损坏权重均拒绝;padding NaN不影响输出,padding动作始终0。
- CPU/CUDA相同输入最大动作差 **1.49e-7**(阈值1e-6),`device="cuda"`正确归一为cuda:0;GPU参数/buffer不变。
- candidate全包manifest hash同时进入物理评估和独立进程报告;最终包再次独立加载,最终manifest hash单独核对。新增“同权重、不同包不能复用证据”回归。
- 已存在发布目录拒绝覆盖;后续任何权重/统计/接口/元数据改动应发新版本。SHA256防意外损坏,不是数字签名;只加载可信来源代码。
最终权重 SHA256:`cdfe459887e7629924df7e6a79fd2900eff184e5b25bc2a26bb585378f671c4a`
最终manifest SHA256:`3106158d2ab214b09360528fb3c994a713435f05f801a8bde5115ebd16079655`
## 7. 工程测试、资源与边界
- 全仓单测 **132 passed, 3 skipped**;skip为保留的可选旧Kit集成测试,不声称重做历史600秒验收;15个外部Lab弃用警告。Ruff与`git diff --check`通过。
- 两次独立只读代码审查共4项发现均修复并验证:耐久尾部内存采样、CUDA设备别名、全包证据身份、文档日志捕获。关键两项有失败回归。
- RTX5080 16GiB单卡;第二轮采集46.85秒、梯度训练134.28秒、最终耐久143.58秒(均不以时长替代质量门槛)。从初始审计到交付约30分钟,低于90分钟预算;两个正式训练尝试,全部本地产物含独立CPU依赖环境约1.2GiB,低于10GB。
- 模拟环境沿用本机Isaac Sim6.0.1路径 / Lab3.0配置,Lab core包6.1.17、Torch2.10.0+cu128;Python与源码hash见`provenance.json`。基线git `8019d6f`,交付工作区有未提交变更;训练源hash、配置hash、数据hash、资产hash及检查点hash均留存,不把dirty源码伪装成已提交版本。
- 外部运行时限制:主仿真进程全部结束后,发现validation_030000遗留一个无GPU context的`carb.tasking6`孤立helper(约1.50GB CPU RSS),已按本次任务PID单独SIGTERM清理并确认退出,记录在`kit_helper_cleanup.json`。这不影响已完成的物理评估或纯Torch冻结推理,但**不宣称Kit批量启动/退出不存在资源残留**;长批次部署仍应使用进程组监督。只读发布目录再次独立加载验证也通过,报告与发布前逐字节一致。
- 不覆盖:Go1/不同DoF泛化、ModSkill latent、接触承重/步态、移动基座、整机协调、噪声/动力学随机化、真实硬件。固定悬空结果不能外推到这些范围。
+7
View File
@@ -0,0 +1,7 @@
# 第二次正式尝试(独立测试前登记)
第一轮完整训练:1,024,000 转移/30,000更新;三个预定validation候选全部失败,不发布。平均误差10.45/11.09/11.90mm;成功率83.0/76.2/72.9%。固定数据中行为策略各噪声组末段均值13.1–14.8mm,说明原行为控制器不足以提供高质量稳态数据。独立test、boundary、stability尚未运行,测试种子未用于任何调参。
使用原定最多两次训练尝试中的第二次:`configs/training/go2_fl_offline_v1_attempt2.json`。范围、样本量、30k更新、IQL超参数、验证选模规则、全部验收门槛、独立test种子**不变**。仅行为收集器增加积分项:I += 8*dt*(q_goal-q)/scale,限幅±0.6,补偿静态重力偏差;噪声档改为[0,.03,.1,.25],收集seed12000,重新完整采集并封存数据。积分只属于数据收集器,不进入发布策略,也不在训练后用解析控制器替换学习策略。
失败第一轮完整数据、checkpoint、日志保留在logs/stage2/v1/;其训练源快照在training_source/(冻结loader后修复了cuda设备规范化,旧checkpoint如需严格恢复须用原快照)。第二轮位置logs/stage2/v1_attempt2/。不能因第一次未过而放宽门槛;第二轮仍不达标则报告未完成。
+27
View File
@@ -0,0 +1,27 @@
# 第二阶段实施与预注册验收 v1
本次第二阶段按用户定义:局部策略离线强化学习预训练与冻结交付,不把 Platform 宏观方案的阶段编号混为一谈。
## 审计
基线 `8019d6f`;开始时 README、train_leg_skill.py 有未提交修改,另有 AGENT.md/test_train_leg_skill.py,均保留。历史日志只有共享 PPO smoke,trained_skill=false。原采样每更新重启 Kit,缺少 optimizer/RNG 恢复、真正离线数据、选模/留出测试及独立冻结发布。v2 描述协议明确尚无通用 adapter,不擅自将旧 token 伪装成 v2。
## 范围与算法(训练前固定)
唯一正式任务为真实 Go2 FL,固定支座悬空、无外载、无接触、无观测噪声,100 Hz 控制/500 Hz 物理。goal 是现有 anchor-frame 足端 link xyz/L,不提供目标关节角给策略。可达目标通过 FK(q_nominal + 0.25*a) 采样,a∈[-0.75,0.75]^3;每目标 2 秒,名义姿态复位,后 0.5 秒判定。不声称 latent 技能库、Go1 泛化、未见形态、步态、承重或真实硬件可用。
收集 20×256×200=1,024,000 条固定转移。行为控制器使用目标关节角 + 位置/速度反馈并混合四档噪声;这是有模型数据收集器,不是发布策略。数据生成后封存 hash;IQL 双 Q、expectile V、优势加权确定性 actor 只读取数据。actor 使用旧 token 中 joint_observation、foot position、goal 的 15 个数,静态绑定 Go2;15 维训练集归一化均值/标准差进入权重,推理不更新。数据末步保留 reset 前 next-state;2 秒为人工截断,继续任务 bootstrap,不把复位观测作为下一状态。
## 预注册阈值、选模及预算
机器配置唯一入口:`configs/training/go2_fl_offline_v1.json`。RTX 5080 单卡;1,024,000 转移,30,000 更新,batch512,最多两次训练尝试,总墙钟预算90分钟/磁盘10GB(不含已有仿真运行时)。预算是上限,不是终止质量门槛。数据20个独立 shard 可恢复,训练每5000步原子保存模型/target/optimizer/RNG;收到 SIGTERM 在当前更新完成后保存。sim 物理状态不声称 bitwise 可恢复;收集按 shard 确定种子重做,梯度训练验证 bitwise resume。
用独立 validation 种子21000的512个目标,比较10k/20k/30k权重,在达到所有 nominal gates 的候选中选末段平均误差最小者;无达标候选则不发布。选定后固定,仅一次独立 test:seed31000 的1024目标 + seed41000 的256边界目标(关节动作空间边/面/角)。报告以 episode 最后50步平均距离为统计单元,成功为该 episode 末段**最大**距离≤15mm;报告置信区间及零动作基线。
Nominal:平均≤8mm、episode均值P95≤15mm、成功率≥95%、相对同目标零动作基线误差降低≥60%。Boundary:平均≤12mm、P95≤20mm,同时硬安全约束通过。所有物理步控制观测:finite、无意外 termination、力矩≤厂家限值(1e-5相对容差)、速度≤厂家限值(1e-5)、关节位置不越硬限位(0.01rad容差)、支座pose漂移≤1e-4。另64实例×120s持续换目标耐久,warmup后设备显存增长≤5%,全部约束零违规。控制周期内5个物理子步的峰值需采样,不仅检查末步。
## 冻结与交付门槛
独立命名 SkillBundle v1(不是 SoftwareSpec v2),只声明现有 leg-token-v1 兼容。包内包含 actor/normalization、接口schema、静态token/物理绑定、资产来源/许可、版本依赖、评估证据、SHA256 manifest与不依赖项目/Kit的推理模块。依赖仅 Python3.12/Torch2.10(CPU 为必须验证平台);禁止覆盖已存在版本。推理 eval + inference_mode + requires_grad=false;无 optimizer、无训练/统计更新路径。参数和buffer调用前后hash一致。包复制到临时独立工作目录,清 PYTHONPATH、隔离进程仅从包加载,覆盖B=1/16/64、错误输入/错绑定/篡改拒绝及稳定推理;不需要训练数据/检查点/资产原路径。
代码/配置/测试/简洁文档入库工作区,完整运行产物位于 logs/stage2/,发布包位于 releases/。只有训练足额完成、独立评估门槛全部达标、独立加载及测试通过才可声明完成;否则保留可恢复产物与缺口报告。
+77
View File
@@ -0,0 +1,77 @@
# Go2 FL Reach 1.0.1 — 完整边界复验与发布
**2026-10-12:已完成真实物理复验并发布。** `logs/stage2/v1_0_1/acceptance.json`中12项检查全部通过,`milestone_complete=true`。该结论仅限固定悬空Go2 FL足端定点到达,不是接触承重、整机、跨形态或实机能力声明。
## 交付
- 只读冻结包:[`releases/go2_fl_reach-1.0.1/`](../releases/go2_fl_reach-1.0.1/)
- 确定性归档:[`go2_fl_reach-1.0.1.tar.gz`](../releases/go2_fl_reach-1.0.1.tar.gz),201,947 bytes
- 外部验收:[`go2_fl_reach-1.0.1.acceptance.json`](../releases/go2_fl_reach-1.0.1.acceptance.json)
- 最终包独立加载:[`go2_fl_reach-1.0.1.independent-load.json`](../releases/go2_fl_reach-1.0.1.independent-load.json)
- 校验入口:[`releases/SHA256SUMS`](../releases/SHA256SUMS)
- 运行证据、源码快照、进程监督记录:`logs/stage2/v1_0_1/`
1.0.0包、归档及原有校验条目原样保留,作为覆盖不全的历史发布;新结果没有覆盖旧报告。
## 不变项与版本管理
复用第二轮的1,024,000条固定数据、完整30,000次训练及原validation-only选择的**20,000步actor**。没有新增正式训练尝试,没有根据边界或测试结果调参、降低阈值或重新选模。发布前再次核对原validation选择规则,结果仍为20k。
`actor.pt`、`selection.json`、`binding.json`、`frozen_policy.py`、`config.json`、`example_tokens.pt`与原candidate逐字节相同;沿用的训练/validation证据也按文件hash核对。actor哈希:
`cdfe459887e7629924df7e6a79fd2900eff184e5b25bc2a26bb585378f671c4a`
训练配置`configs/training/go2_fl_offline_v1_attempt2.json`完全未改。新增`--release-version`和包内`release.json`独立声明发布版本并绑定训练配置hash;`config.json.release_version=1.0.0`是历史训练配置中的默认值,**实际发布版本以manifest/release.json一致的1.0.1为准**。错误版本格式、配置身份不匹配或finalize时修改被测candidate版本会拒绝。
## 真实仿真结果
同一候选包manifest下重新执行boundary、test、stability,使用Isaac/PhysX真实环境而非单元测试fixture;每个模式独立启动headless Kit。由于版本元数据改变了包身份,三个模式均重跑,未替换旧报告的hash字段。
| 复验 | 样本及覆盖 | 均值 | episode均值P95 | 成功率 |
|---|---|---:|---:|---:|
| 完整边界,seed41000 | 256目标,6个面/12条边/8个角 | **2.308mm** | **3.286mm** | **100%** |
| 名义测试,seed31000–31003 | 1024目标 | **2.184mm** | **4.532mm** | **100%** |
- 边界每种轴/符号类别9或10个目标,保存送入FK的实际动作并重新计算覆盖率;是全部26种边界**类别**覆盖,不是穷尽连续边界。
- 边界门槛仍为均值≤12mm、P95≤20mm及原物理安全门槛;名义门槛仍为≤8mm/≤15mm、成功≥95%、相对零动作基线改善≥60%。成功要求episode最后50个控制步全部≤15mm。
- 名义结果与旧权重评估相同,相对零动作基线误差下降95.881%。这次复用相同holdout种子,**不声称新增未见测试集**;边界采样协议改变,不能将新旧边界均值差当作训练收益。
- 边界峰值归一化torque/velocity为0.406347/0.786603;所有本次物理评估硬限位超出、支座漂移、invalid states均为0,无意外termination/reset。
## 耐久与运行时
- 64环境×120秒,12,000控制步、768,000环境控制步;另200步warmup,合计61,000批物理子步快照。目标每2秒更换、不中途复位。
- 显存初始3,525,246,976 bytes,峰值3,554,869,248 bytes,增长**0.8403%≤5%**,包含尾部采样。
- 峰值normalized torque=0.663709、velocity=1.000000834≤1.00001。速度触及上限,**不宣称限速安全裕量**;torque是仿真applied torque,不是实测电机力矩。
- 三个模式主进程退出及日志扫描均通过。监督器只收养、终止并回收本任务遗留的Hub/telemetry helper,三个`*_process.json`均`cleanup.clean=true`;不是对Kit原生退出无残留的保证。
- 边界/名义/耐久含启动与清理墙钟约21.66/36.05/182.33秒。仍为单GPU运行,没有并行Kit任务。
## 恢复、工程与冻结加载
- 两个新的独立GPU进程从真实5k检查点恢复到5001,**112个tensor**及非tensor结构、optimizer、CPU/CUDA RNG逐项一致。只用于恢复复验,不将这些probe作为训练完成或重新选模证据。
- 全仓pytest **172 passed, 3 skipped**,Ruff、`git diff --check`通过。3个skip仍是历史可选Kit集成测试,不声称重做旧600秒验收。新增发布版本与训练身份解耦的回归。
- candidate、暂存最终包、只读正式包均通过独立CPU解释器`python -I`包-only加载;B=1/16/64/256的确定性调用、输入/参数/buffer不变、错误绑定与损坏拒绝通过。无需Isaac、训练文件、项目导入或资产路径。
- 新CPU/CUDA parity最大动作差**4.1723e-7<1e-6**;400次GPU重复调用确定性,参数/buffer不变,`cuda`别名解析为cuda:0。
- 工程/恢复/物理/独立加载等语义门槛通过后,暂存最终包再次独立加载、统一总验收通过,才原子发布。正式目录随后设为只读并再次独立加载、总验收。
## 完整性与复核
最终manifest SHA256:
`b685088e675fc624744a239644ce59766efab6bbb8f4790b4783c16f0ed100fc`
归档 SHA256:
`90168f81e33abbe1f5ac49a71cc9447b61dea6db9bbb886878d15134695c9dc7`
```bash
unset CONDA_PREFIX VIRTUAL_ENV
export PYTHONPATH="$PWD/source/SkillFactory"
/home/cen/IsaacLab/isaaclab.sh -p scripts/verify_stage2_delivery.py \
--config configs/training/go2_fl_offline_v1_attempt2.json \
--run logs/stage2/v1_0_1 --release releases/go2_fl_reach-1.0.1 \
--output logs/stage2/recheck_1_0_1/acceptance.json
python3 scripts/verify_frozen_leg.py --package releases/go2_fl_reach-1.0.1 \
--python logs/stage2/inference_venv/bin/python \
--output logs/stage2/recheck_1_0_1/standalone.json
(cd releases && sha256sum -c SHA256SUMS)
```
归档和SHA256用于完整性而非数字签名;只运行可信包代码。工作区仍为带有既有修改的dirty状态,未创建git提交,来源与hash见新run的`provenance.json`。推理延迟测量/性能优化没有包含在本次任务中。
+42
View File
@@ -0,0 +1,42 @@
# 第二阶段发布可靠性修复
## 状态
**后续已完成[1.0.1完整物理复验及发布](stage2_release_1_0_1.md),沿用原权重且未重训。** 以下保留修复阶段的历史记录:当时三项问题已修复并通过失败路径回归,但尚未物理复验/发布;1.0.0文件与原校验条目始终未改。新版本的证据单独位于`logs/stage2/v1_0_1/`,SHA256SUMS仅追加新版本条目。
## 修复
1. **P1 发布门槛遗漏**
- 新增`scripts/stage2_acceptance.py`,发布器和总验收器共用数据、训练、选模、恢复、工程、物理、独立加载检查;不再各维护一套不同门槛。
- 工程报告及完整learner/optimizer/RNG恢复报告必须存在、解析成功且通过。检查工程日志hash,重算episode统计,校验物理时长/样本数、边界覆盖及整个被测包的manifest身份。
- 先验证全部语义门槛,再在同文件系统的临时目录构建最终包;使用必需的`--python`独立验证**最终manifest字节**,再复用总验收逻辑检查,最后原子rename公开。失败不留下公开accepted目录;同版本禁止覆盖。
- 新版总验收亦先写running,缺失/损坏证据会覆盖旧成功报告为failed;可用`--output`保存历史报告。
2. **P2 边界覆盖不全**
- 新增`scripts/stage2_boundary.py`:枚举三轴的全部26种非零边界签名,6个面内点类别、12条边内点类别、8个角。
- 循环分配每个轴/符号组合,非固定坐标严格位于内部,固定seed可复现;256环境下每种组合9或10个目标。少于26环境直接拒绝,不静默缩小覆盖范围。
- 从实际送往FK的动作记录`boundary_actions`和`boundary_coverage`,不是只记录预期采样计划。评估和发布都要求覆盖完整;旧协议证据拒绝通过。
- 这里的“完整”指全部边界**类别**均被采样,不声称穷尽连续边界上的所有点。
3. **P2 失败复验残留旧成功报告**
- `verify_frozen_leg.py`启动立即原子写`passed=false,status=running`和新run_id。
- 复制失败、解释器缺失、子进程非零、畸形/否定输出、可捕获中断等写failed并重新抛错,保留子进程returncode/stdout/stderr。突然kill保留的running同样不能通过门槛。
- 只有校验成功、输入包身份未变化、临时目录清理完成才写completed。隔离子进程仍只依赖stdlib/Torch,不需要新验收模块。
## 修复阶段验证(发布复验之前)
- 全仓pytest:**167 passed, 3 skipped**(新增35项,历史可选Kit集成仍skip);Ruff、`git diff --check`通过。
- 回归包含:工程/恢复分别失败及同时失败、缺失/畸形JSON;旧边界证据、缺日志、物理步数不足、不同manifest复用相同权重;暂存复制/加载失败不发布;全部门槛满足后的原子发布及拒绝覆盖;3种batch ×3个seed的全部轴/符号覆盖;旧成功被失败复验覆盖、成功路径及包变化拒绝。
- 发布事务单元测试中的物理数据是显式合成fixture,最终独立加载调用在该单元测试中mock;**这些不是新物理验收**。
- 另用真实独立CPU解释器验证原只读1.0.0包:通过。再在临时副本注入权重损坏,对预先填入成功的报告复验:实际子进程失败、CLI退出1,报告变为`passed=false,status=failed`。
- `releases/SHA256SUMS`全部匹配,训练源码身份绑定的三个文件未改动。
证据:`logs/stage2/reliability_fix/{pytest.log,ruff.log,readonly_standalone.json,failed_recheck.json,historical_delivery_recheck.json}`。
## 历史产物的处理与原待办
新版检查对旧1.0.0报告预期返回非零:
- 训练、恢复、选模、数据、名义测试、耐久、包完整性仍通过已有证据核对。
- **边界不通过**:旧报告实际为2/6个面、4/12条边、8/8个角,没有新协议的完整采样证据。
- 历史两份独立加载报告没有新版`status=completed`生命周期字段,不能直接复用;本次真实CPU复验结果单独保留,未篡改历史报告。
上述原待办已在1.0.1完成:不改变权重/阈值、不利用测试选模,重新执行完整边界、名义测试、耐久、恢复一致性、工程及独立加载,原子发布新版本。不要修改旧1.0.0内容或仅手工补字段让旧证据“通过”。推理延迟测量和基于瓶颈的性能优化尚未进行。
+100
View File
@@ -0,0 +1,100 @@
# 第二阶段:离线单腿 Reach 与冻结调用
范围、预算、预注册门槛:[stage2_plan.md](stage2_plan.md)。第二次尝试理由:[stage2_attempt2.md](stage2_attempt2.md)。与通用描述规范的关系:继续复用现有 leg-token-v1,不声称已实现 SoftwareSpec v2 通用 adapter。
## 重现训练
先按项目 README 构建 `assets/generated/legs/go2_fl`。Isaac Sim 6.0.1 / Lab 3.0.0、Python3.12、Torch2.10 CUDA;收集和评估需要 Kit,梯度训练不需要 Kit。
```bash
unset CONDA_PREFIX VIRTUAL_ENV
export PYTHONPATH="$PWD/source/SkillFactory"
export PYTHONUNBUFFERED=1
LAB=/home/cen/IsaacLab/isaaclab.sh
CFG=configs/training/go2_fl_offline_v1_attempt2.json
RUN=logs/stage2/reproduction # 使用新目录;不要覆盖交付证据
VERSION=1.0.2 # 重现示例:选择尚未占用的版本;当前已交付1.0.1
mkdir -p "$RUN"
set -o pipefail
"$LAB" -p scripts/stage2_sim.py --config "$CFG" --mode collect --output "$RUN/data" --device cuda:0 2>&1 | tee "$RUN/collect.log"
"$LAB" -p scripts/train_offline_leg.py --config "$CFG" --data "$RUN/data" --output "$RUN/train" 2>&1 | tee "$RUN/train.log"
```
数据收集20个确定性种子shard;中断后重跑收集命令只重新生成未完成shard。固定数据封存后训练绝不调用仿真器。训练每5000更新保存actor/Q/V/target/optimizer/RNG;SIGINT/SIGTERM在当前更新后原子落盘。磁盘/硬崩溃可退回最近一次完整checkpoint。恢复必须使用相同config、dataset及训练源hash:
```bash
"$LAB" -p scripts/train_offline_leg.py --config "$CFG" --data "$RUN/data" \
--output "$RUN/train" --resume "$RUN/train/latest.pt"
```
`--stop-after 5000` 是中断演练,不是训练完成。完整必须达到30000更新且训练result.training_complete=true;它仍不等于任务验收通过。所有JSON报告以passed=false开始,shutdown失败不得标成功。
## 选模、独立评估与发布
**发布可靠性修订:** 历史1.0.0的边界报告只有2/6个面、4/12条边、8/8个角,不能作为新门槛的完整覆盖证据。修订后的采样按26种轴/符号组合循环,每批至少26环境,并保存实际动作及覆盖计数。原训练/权重未更改;完整边界、名义测试、耐久等已重新执行并发布1.0.1,见[发布报告](stage2_release_1_0_1.md)。旧版问题见[修复记录](stage2_reliability_fix.md)。
```bash
for STEP in 010000 020000 030000; do
"$LAB" -p scripts/stage2_sim.py --config "$CFG" --mode validation \
--weights "$RUN/train/step_$STEP.pt" --output "$RUN/validation_$STEP" --device cuda:0 2>&1 | tee "$RUN/validation_$STEP.log"
# 某候选未达标退出非零,可记录后继续其它预定候选;运行异常必须检查日志。
done
"$LAB" -p scripts/export_leg_bundle.py --config "$CFG" --run "$RUN" --output "$RUN/candidate" \
--release-version "$VERSION"
for MODE in test boundary stability; do
"$LAB" -p scripts/stage2_sim.py --config "$CFG" --mode "$MODE" \
--package "$RUN/candidate" --output "$RUN/$MODE" --device cuda:0 2>&1 | tee "$RUN/$MODE.log"
done
```
导出先检查完整训练与三份validation证据,并仅在通过门槛者中按平均误差选模。未达标不导出。candidate已经冻结,但不是accepted release。test种子不能用于调参或选模;任一独立门槛不通过,停止发布。每步检查退出码、result与日志的`[Error]/[Fatal]/buffer overflow`,不能只看奖励。
独立CPU验证(在普通Python新venv中,不用Kit):
```bash
python3 -m venv /tmp/skill-inference
/tmp/skill-inference/bin/python -m pip install torch==2.10.0 --index-url https://download.pytorch.org/whl/cpu
python3 scripts/verify_frozen_leg.py --package "$RUN/candidate" \
--python /tmp/skill-inference/bin/python --output "$RUN/standalone.json"
# 发布前必须已有本次真实工程测试、恢复一致性证据,不能复制历史passed标志:
# $RUN/engineering_checks.json、$RUN/resume_exact.json(字段要求见下文)。
# candidate的release.json/manifest版本已固定;不能在finalize时改版本或复用旧包测试hash。
"$LAB" -p scripts/export_leg_bundle.py --config "$CFG" --run "$RUN" --candidate "$RUN/candidate" \
--python /tmp/skill-inference/bin/python --output "releases/go2_fl_reach-$VERSION"
python3 scripts/verify_frozen_leg.py --package "releases/go2_fl_reach-$VERSION" \
--python /tmp/skill-inference/bin/python --output "$RUN/release_standalone.json"
```
`--release-version`只在candidate导出时定义发布元数据,不改变已hash锁定的训练配置。包内`release.json`绑定训练配置hash并与manifest版本一致;旧配置的`release_version`保留为历史默认值,不能通过修改它绕过训练身份锁定。省略参数时保持旧配置默认版本。仅换版本也改变candidate manifest,因此所有包级物理/独立加载证据必须对应新包。
发布器与交付核验器共享`stage2_acceptance.py`:数据/训练/选模、工程测试、恢复一致性、物理指标及覆盖、独立加载全部通过后,才在同文件系统临时目录构建最终包,并用`--python`再次独立验证最终manifest对应的字节;通过后原子rename为公开accepted目录。中途失败会清理临时包,不会遗留公开accepted目录。
工程报告必须`passed=true`、`git_diff_check=true`、`review_blockers_remaining=0`,并有与实际`pytest.log`/`ruff.log`匹配的`pytest_log_sha256`/`ruff_log_sha256`。恢复报告必须`passed=true`且`full_learner_optimizer_rng_exact=true`,均为强制门槛,缺失/损坏/失败均拒绝发布。
验证工具启动即原子写`passed=false,status=running`;异常(复制、解释器/子进程失败、畸形输出、中断)原子写`passed=false,status=failed`并非零退出,不复用旧成功报告。成功状态为`completed`,附独立run_id。突然kill留下的running也不能通过发布门槛。
验证工具把**仅包和验证脚本**复制到全新临时工作目录,清掉Python/Isaac环境变量,以`python -I`启动;检查B=1/16/64/256,400次确定性调用、参数/buffer/输入不变、错输入/绑定/篡改拒绝。后5槽始终0。模拟评估与独立加载都绑定**整个candidate manifest hash**,不能只凭同一actor复用不同包的证据。发布目录拒绝覆盖;后续任何修改必须新版本。
重新核验最新1.0.1时使用新报告路径保留交付证据,应全部通过。若改为历史1.0.0与`v1_attempt2`运行目录,则因旧边界覆盖及旧加载报告生命周期字段缺失而返回非零;不要篡改历史报告。全部版本文件SHA256均应匹配:
```bash
"$LAB" -p scripts/verify_stage2_delivery.py \
--config configs/training/go2_fl_offline_v1_attempt2.json \
--run logs/stage2/v1_0_1 --release releases/go2_fl_reach-1.0.1 \
--output logs/stage2/recheck_1_0_1/acceptance.json
(cd releases && sha256sum -c SHA256SUMS)
```
发布器和最终核验器都会读取`resume_exact.json`与`engineering_checks.json`;历史交付已包含真实恢复/测试证据,但不能据此替代新代码的复验。新训练不能复用这些结果,须重做恢复一致性、pytest、Ruff并记录自己的报告。
实际推理示例见包内README;输入实时token,`act`返回归一化PD目标偏移,`position_targets`返回软限位后的声明顺序rad目标。接收方必须提供自己的资产manifest hash,不应盲信包内示例。无物理仿真依赖,无训练临时文件/optimizer;Torch可发出缺少可选NumPy的警告,但推理不使用NumPy。
## 测试与证据
```bash
PYTEST_DISABLE_PLUGIN_AUTOLOAD=1 "$LAB" -p -m pytest tests -q
"$LAB" -p -m ruff check source/SkillFactory scripts tests
```
`tests/test_offline_leg.py`覆盖IQL更新、连续vs恢复完全一致、非有限loss、冻结状态、padding、hash拒绝及耐久尾部内存增长回归。`tests/test_stage2_release.py`覆盖强制门槛失败/缺失/损坏、原子发布及最终包加载失败、全部26种边界组合覆盖、旧成功报告被失败复验覆盖等路径。测试之外保留真实大数据训练恢复和完整独立仿真评估证据。历史第一周仿真集成三项默认skip,不声称此次重做旧600秒验收。
+15
View File
@@ -0,0 +1,15 @@
# 发布状态说明
**固定悬空Reach当前推荐版本为`go2_fl_reach-1.0.1/`及同名归档**:完整边界6面/12边/8角、名义测试、耐久、恢复、工程和独立加载复验通过,12项总验收全部通过。详见[1.0.1发布报告](../docs/stage2_release_1_0_1.md)。
`go2_fl_reach-1.0.0/`及同名归档、校验文件保留原样,作为历史产物。
后续复核发现发布门槛遗漏工程/恢复验证、边界只覆盖2/6个面及4/12条边(8角完整)、失败独立复验可能残留成功报告。代码现已修复,并在1.0.1使用同一权重完成真实物理复验;新证据没有回填旧包。
因此旧包内的`status=accepted`及旧验收JSON仅代表当时的检查结果,**不代表通过修正后的完整发布门槛**。既有30,000步训练及名义任务评估数据没有被改写;能力仍限固定悬空Go2 FL。
详见[`docs/stage2_reliability_fix.md`](../docs/stage2_reliability_fix.md)。
## 受载站立:带偏差豁免的1.0.0
`go2_fl_loaded_stand-1.0.0/`及同名归档是独立任务的冻结发布。用户明确批准当前候选在120秒耐久中30.62/30.78mm漂移相对30mm门槛的特例;标记`accepted_with_deviation`,**不是原始门槛全部通过**,不改变全局门槛。其余门槛、隔离加载和归档核验通过,原始失败证据完整保留。详见[发布报告](../docs/loaded_stand_release_1_0_0.md)及包内`DEVIATIONS.md`;不可把该豁免用于其它权重或试验。
+13
View File
@@ -0,0 +1,13 @@
f7d426fa23ba3ab72bf5a4f1a87845c0cfa8b6167c631061a7d00823c7dcd05b go2_fl_reach-1.0.0.tar.gz
3106158d2ab214b09360528fb3c994a713435f05f801a8bde5115ebd16079655 go2_fl_reach-1.0.0/manifest.json
fce757d5b0ee310e9660907bcaa1c18d71d0c01095552c99fbf25d171c1860b7 go2_fl_reach-1.0.0.acceptance.json
2c6574c098682f93be7f5923cd63275b07ea6959056b24ce0947d26ce4c6977e go2_fl_reach-1.0.0.independent-load.json
90168f81e33abbe1f5ac49a71cc9447b61dea6db9bbb886878d15134695c9dc7 go2_fl_reach-1.0.1.tar.gz
b685088e675fc624744a239644ce59766efab6bbb8f4790b4783c16f0ed100fc go2_fl_reach-1.0.1/manifest.json
6ee59208035250a1b7bfefb34d3cd157316b1522a893dfbb5cc554f485b38d2e go2_fl_reach-1.0.1.acceptance.json
5e722e39c1d366064010fc8932f2e48edf9be5b291d5f67afc36da7df0131d0b go2_fl_reach-1.0.1.independent-load.json
ad59972e041f079cc2584923aaa15798857e5619488ba76b69835b558956bd76 go2_fl_loaded_stand-1.0.0.tar.gz
27e5e56908a0a8585a13913f27663b0dd3cdb7270614a0a05355334de4093e98 go2_fl_loaded_stand-1.0.0/manifest.json
4a8fc617770b74dead5d331259872571094b203543dbf2b070028f9e3de27b77 go2_fl_loaded_stand-1.0.0.acceptance.json
d50c4e8141c7fa4a4773246568f14e6501829f21959218d9b8c64744ad75fab0 go2_fl_loaded_stand-1.0.0.independent-load.json
760abb6f7fbf624f9a5f4c3a8526118e4d386e43b60809838e9de0c00b013181 go2_fl_loaded_stand-1.0.0.archive-load.json
@@ -0,0 +1,256 @@
{
"passed": true,
"status": "completed",
"accepted_skill": true,
"acceptance_basis": "operator_waiver",
"unwaived_required_physics_passed": false,
"waiver_sha256": "3c8c03efcedc3c6eb2a0a8ca8fdae2629834473fb122b2ce5bbcc39161344d1e",
"waiver_receipts": [
{
"accepted_by_operator_waiver": true,
"raw_gate_passed": false,
"gate": "stability.foot_drift",
"original_limit_m": 0.03,
"observed_failures": [
{
"environment": 6,
"max_foot_drift_m": 0.030620924223339147
},
{
"environment": 19,
"max_foot_drift_m": 0.03078258298377699
}
]
}
],
"waiver": {
"format": "skillfactory.standing-deviation.v1",
"authorization": "explicit_user_instruction",
"operator_statement": "\u8fd9\u4e2a120\u79d2\u8010\u4e45\u7684\u98d8\u9038\u8d85\u8fc7\u7684\u5f88\u5c0f\uff0c\u53ef\u4ee5\u89c6\u4e3a\u901a\u8fc7",
"recorded_at": "2026-10-12T16:58:57.445471+08:00",
"approved": true,
"scope": "exact_candidate_and_trial_only",
"waived_gate": "stability.foot_drift",
"original_limit_m": 0.03,
"candidate_manifest_sha256": "7cb478a3b17aaf307d2d6e3c2b49cc25aa9af1ee9d118dd731b52f618c27e628",
"actor_sha256": "1384b5aa780cded1f4ed21b617ff8c883a5f82826c4cb4c9415b75f315b1e664",
"checkpoint_sha256": "f2197339dea3622ed92b82fdddff3400f2a82f3ca4d4fd3be6c72b45f1a77704",
"duration_s": 120.0,
"num_envs": 26,
"physical_samples": 60000,
"observed_failures": [
{
"environment": 6,
"max_foot_drift_m": 0.030620924223339147
},
{
"environment": 19,
"max_foot_drift_m": 0.03078258298377699
}
],
"evidence_sha256": {
"stability/result.json": "95a0f18edac40ece69430041b5083a011fe612a4770e899575ca985ba496e0b4",
"stability/metadata.json": "4c26c6fc3f2eb59b0fe6f2a5d1962c283f2da8e2c4047a33929f6643b17ab447",
"stability/physics_trace.pt": "3b35f32f71e87c812a24ae93101f05c917707691fea15ee523cca161846fe498",
"stability/memory.json": "3360e264f6062eff14fd1a015e670c26bdc1f84d1b7acc9dd568c35fd63868c5",
"stability_process.json": "e2af5a4fd4341dcb4ccecdef54dfa9c6404133a54cf2d7d05179056095ddd26d",
"stability.log": "e7678de92de394288d0fa5e570de090cb323bbd23411a49a7390f6ef30c6f380"
},
"not_authorized": [
"changing_original_results",
"changing_global_threshold",
"waiving_other_gates",
"new_weights_or_trials",
"hardware_safety_claims"
]
},
"version": "1.0.0",
"scope": "guided single Go2 FL, PGS, synthetic3kg payload, single-axis validated loads; not hardware",
"manifest_sha256": "27e5e56908a0a8585a13913f27663b0dd3cdb7270614a0a05355334de4093e98",
"candidate_manifest_sha256": "7cb478a3b17aaf307d2d6e3c2b49cc25aa9af1ee9d118dd731b52f618c27e628",
"archive_sha256": "ad59972e041f079cc2584923aaa15798857e5619488ba76b69835b558956bd76",
"core_files": {
"ASSET_LICENSE": "84aac59fd3246e3ddc49d1387644e8fcf43b0def4f0b9fe687f372e90446df2d",
"README.md": "2c8cb0ea4f057b5c07600f2054aa58eed35b5e45e253ce7eec9bc1a532fc57ea",
"actor.pt": "1384b5aa780cded1f4ed21b617ff8c883a5f82826c4cb4c9415b75f315b1e664",
"binding.json": "22cd42a44f29913b2f0bd04e685e77acdcc1ef7b0218f62733dc401d55b6d710",
"example.pt": "f21e71bbc4b0e862a672d13daf67c31971f365426dd98662fc658ede6916c798",
"frozen_standing.py": "c19026845ea5ec98b21ed641cd294cccd502cf73907c08f0fe71a9a63067d8c6",
"requirements.txt": "812ea7d306f227902500eb76d816fe9f65f75d201edc6e766118ed68d3fdfe0d",
"selection.json": "276302e9374337aed8d4d3ea35408c4a0dc9971ea37bcecda8ccc5208b6a296a",
"task.json": "bb2671ba1f1750baf6f0ec3ebb2f555d60ee252358867aa378fc85e07790c99a",
"training.json": "443187246b6a600a11566eb1f1d3dca260f7c6a64684a847af7577afc4e521c9"
},
"trials": {
"full": {
"passed": true,
"accepted_by_operator_waiver": false,
"gates": {
"finite_complete_trace": true,
"floating_root": true,
"pgs_compatible": true,
"no_autoreset": true,
"no_fall": true,
"applied_load_matches_protocol": true,
"quaternion_norm": true,
"no_vertical_guide": true,
"height": true,
"tilt": true,
"foot_drift": true,
"contact": true,
"vertical_balance": true,
"limits": true,
"sensor_static_force": true,
"sensor_static_torque": true,
"indexed_reset": true,
"timeout_terminal_reset": true,
"bad_action_atomic_rejection": true
},
"mean_height_error_m": 0.0010973225607199789,
"cpu_cuda_max_error": 1.7881393432617188e-07,
"physical_samples": 7000,
"max_height_error_m": 0.011397260259590258,
"max_tilt_rad": 0.039559598305326425,
"max_foot_drift_m": 0.004637469188639254,
"min_contact_fraction": 1.0,
"sensor_force_error_N": 0.029688779830932432,
"sensor_torque_error_Nm": 0.001858974476667002,
"memory_growth_fraction": null
},
"holdout": {
"passed": true,
"accepted_by_operator_waiver": false,
"gates": {
"finite_complete_trace": true,
"floating_root": true,
"pgs_compatible": true,
"no_autoreset": true,
"no_fall": true,
"applied_load_matches_protocol": true,
"quaternion_norm": true,
"no_vertical_guide": true,
"height": true,
"tilt": true,
"foot_drift": true,
"contact": true,
"vertical_balance": true,
"limits": true
},
"mean_height_error_m": 0.0009209314090133883,
"cpu_cuda_max_error": 1.7881393432617188e-07,
"physical_samples": 7000,
"max_height_error_m": 0.01069288236995708,
"max_tilt_rad": 0.03849515058506419,
"max_foot_drift_m": 0.004441969886462753,
"min_contact_fraction": 1.0,
"sensor_force_error_N": 0.11419118010997775,
"sensor_torque_error_Nm": 0.007193774832685795,
"memory_growth_fraction": null
},
"hold": {
"passed": false,
"accepted_by_operator_waiver": false,
"gates": {
"finite_complete_trace": true,
"floating_root": true,
"pgs_compatible": true,
"no_autoreset": true,
"no_fall": true,
"applied_load_matches_protocol": true,
"quaternion_norm": true,
"no_vertical_guide": true,
"height": false,
"tilt": true,
"foot_drift": true,
"contact": true,
"vertical_balance": true,
"limits": true
},
"mean_height_error_m": 0.06637654078181726,
"cpu_cuda_max_error": 0.0,
"physical_samples": 7000,
"max_height_error_m": 0.09780764861007701,
"max_tilt_rad": 0.05186780450752031,
"max_foot_drift_m": 0.0033128396003425657,
"min_contact_fraction": 1.0,
"sensor_force_error_N": 0.12399852900206909,
"sensor_torque_error_Nm": 0.005068768669592938,
"memory_growth_fraction": null
},
"stability": {
"passed": false,
"accepted_by_operator_waiver": true,
"gates": {
"finite_complete_trace": true,
"floating_root": true,
"pgs_compatible": true,
"no_autoreset": true,
"no_fall": true,
"applied_load_matches_protocol": true,
"quaternion_norm": true,
"no_vertical_guide": true,
"height": true,
"tilt": true,
"foot_drift": false,
"contact": true,
"vertical_balance": true,
"limits": true,
"sensor_static_force": true,
"sensor_static_torque": true,
"memory": true
},
"mean_height_error_m": 0.0017030573373548721,
"cpu_cuda_max_error": 2.384185791015625e-07,
"physical_samples": 60000,
"max_height_error_m": 0.020278918575248828,
"max_tilt_rad": 0.057941022266468155,
"max_foot_drift_m": 0.03078258298377699,
"min_contact_fraction": 1.0,
"sensor_force_error_N": 0.02974697828292827,
"sensor_torque_error_Nm": 0.0018565384255236957,
"memory_growth_fraction": 0.0
},
"exploratory": {
"passed": true,
"accepted_by_operator_waiver": false,
"gates": {
"finite_complete_trace": true,
"floating_root": true,
"pgs_compatible": true,
"no_autoreset": true,
"no_fall": true,
"applied_load_matches_protocol": true,
"quaternion_norm": true,
"no_vertical_guide": true,
"height": true,
"tilt": true,
"foot_drift": true,
"contact": true,
"vertical_balance": true,
"limits": true
},
"mean_height_error_m": 0.001673169419981864,
"cpu_cuda_max_error": 1.7881393432617188e-07,
"physical_samples": 7000,
"max_height_error_m": 0.019595432113609423,
"max_tilt_rad": 0.05547213523615746,
"max_foot_drift_m": 0.007914748932079697,
"min_contact_fraction": 1.0,
"sensor_force_error_N": 0.05928855037689118,
"sensor_torque_error_Nm": 0.003488774794082272,
"memory_growth_fraction": null
}
},
"holdout_improvement_fraction": 0.9861256492404368,
"independent_candidate": true,
"independent_staged": true,
"independent_final": true,
"independent_archive": true,
"raw_trace_reduction": true,
"training": {
"updates": 2048,
"transitions": 16777216,
"selected_update": 1024,
"checkpoint_sha256": "f2197339dea3622ed92b82fdddff3400f2a82f3ca4d4fd3be6c72b45f1a77704"
}
}
@@ -0,0 +1,24 @@
{
"passed": true,
"package_only": true,
"isolated": true,
"device": "cpu",
"python": "3.12.3 (main, Aug 31 2026, 10:18:26) [GCC 13.3.0]",
"torch": "2.10.0+cpu",
"manifest_sha256": "27e5e56908a0a8585a13913f27663b0dd3cdb7270614a0a05355334de4093e98",
"actor_sha256": "1384b5aa780cded1f4ed21b617ff8c883a5f82826c4cb4c9415b75f315b1e664",
"forbidden_imports": [],
"deterministic_batches": [
1,
16,
64,
256
],
"calls_per_batch": 100,
"input_state_unchanged": true,
"requires_grad": false,
"invalid_inputs_binding_corruption_rejected": true,
"status": "completed",
"package": "/tmp/standing-archive-0_sw9i7f/go2_fl_loaded_stand-1.0.0",
"stderr": "/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/stage2/inference_venv/lib/python3.12/site-packages/torch/_subclasses/functional_tensor.py:283: UserWarning: Failed to initialize NumPy: No module named 'numpy' (Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:84.)\n cpu = _conversion_method_template(device=torch.device(\"cpu\"))\n"
}
@@ -0,0 +1,24 @@
{
"passed": true,
"package_only": true,
"isolated": true,
"device": "cpu",
"python": "3.12.3 (main, Aug 31 2026, 10:18:26) [GCC 13.3.0]",
"torch": "2.10.0+cpu",
"manifest_sha256": "27e5e56908a0a8585a13913f27663b0dd3cdb7270614a0a05355334de4093e98",
"actor_sha256": "1384b5aa780cded1f4ed21b617ff8c883a5f82826c4cb4c9415b75f315b1e664",
"forbidden_imports": [],
"deterministic_batches": [
1,
16,
64,
256
],
"calls_per_batch": 100,
"input_state_unchanged": true,
"requires_grad": false,
"invalid_inputs_binding_corruption_rejected": true,
"status": "completed",
"package": "/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/releases/go2_fl_loaded_stand-1.0.0",
"stderr": "/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/stage2/inference_venv/lib/python3.12/site-packages/torch/_subclasses/functional_tensor.py:283: UserWarning: Failed to initialize NumPy: No module named 'numpy' (Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:84.)\n cpu = _conversion_method_template(device=torch.device(\"cpu\"))\n"
}
@@ -0,0 +1,24 @@
{
"passed": true,
"package_only": true,
"isolated": true,
"device": "cpu",
"python": "3.12.3 (main, Aug 31 2026, 10:18:26) [GCC 13.3.0]",
"torch": "2.10.0+cpu",
"manifest_sha256": "27e5e56908a0a8585a13913f27663b0dd3cdb7270614a0a05355334de4093e98",
"actor_sha256": "1384b5aa780cded1f4ed21b617ff8c883a5f82826c4cb4c9415b75f315b1e664",
"forbidden_imports": [],
"deterministic_batches": [
1,
16,
64,
256
],
"calls_per_batch": 100,
"input_state_unchanged": true,
"requires_grad": false,
"invalid_inputs_binding_corruption_rejected": true,
"status": "completed",
"package": "/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/releases/.go2_fl_loaded_stand-1.0.0.stage-oc48qrml",
"stderr": "/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/stage2/inference_venv/lib/python3.12/site-packages/torch/_subclasses/functional_tensor.py:283: UserWarning: Failed to initialize NumPy: No module named 'numpy' (Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:84.)\n cpu = _conversion_method_template(device=torch.device(\"cpu\"))\n"
}
Binary file not shown.
@@ -0,0 +1,29 @@
BSD 3-Clause License
Copyright (c) 2016-2022 HangZhou YuShu TECHNOLOGY CO.,LTD. ("Unitree Robotics")
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are met:
* Redistributions of source code must retain the above copyright notice, this
list of conditions and the following disclaimer.
* Redistributions in binary form must reproduce the above copyright notice,
this list of conditions and the following disclaimer in the documentation
and/or other materials provided with the distribution.
* Neither the name of the copyright holder nor the names of its
contributors may be used to endorse or promote products derived from
this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
@@ -0,0 +1,3 @@
# Accepted with an operator-approved deviation
The original120s stability foot-drift gate FAILED: env6=30.620924mm, env19=30.782583mm versus30mm. Raw results and the original threshold are unchanged. The operator accepted this exact candidate and recorded trial only; see evidence/operator_waiver.json and manifest waiver_sha256. This is NOT a general31mm threshold, a new physics pass, or a hardware safety approval. No other failed gate is waived; no new weights or trial may reuse this waiver. The archive and final directory are independently loaded after eligibility checks.
@@ -0,0 +1,17 @@
# Go2 FL guided loaded stand 1.0.0
Inference requires Python3.12/Torch2.10.0 only, no Kit/project/training checkpoint. Load FrozenStanding(directory, manifest_sha256=<trusted manifest hash>, task_sha256=<deployed task hash>, support_sha256=<deployed support manifest hash>, device='cpu'). act(observation) returns float32[B,3] bounded normalized position actions; position_targets returns rad. See example.pt and binding.json. Caller supplies six exact observation fields with mixed frames, unit xyzw quaternion, actual last applied actions, true six-axis mount wrench, 100Hz sampling, and all driver limits.
Scope: floating 3kg synthetic payload, Go2 FL, PGS, XY/rotation guidance, zero vertical assistance; not unsupported balance, full quadruped, arbitrary mechanics, general v2 adapter or hardware certification. Source missing inertias/imported assumptions are disclosed in binding support_manifest; no source mechanical changes. The asset itself is externally bound, not required for inference or bundled as a turnkey simulator.
Checkpoint1024 was selected by validation only. Validation is not holdout or acceptance. Only an accepted manifest together with its bound acceptance evidence establishes a release. Two-axis combinations are exploratory and outside validated simulation scope. SHA256 is not a signature.
## Exact observation ABI (all float32, same device, nonempty batch B)
- joint_state[B,3,3]: declared joint order; q(rad), dq(rad/s), previous normalized action.
- base_state[B,13]: payload mounting-origin position minus environment origin (m); world orientation xyzw; world linear velocity(m/s); world angular velocity(rad/s). The validated ground/environment origins have world z=0. Not root-frame velocities.
- mount_wrench[B,6]: [Fx,Fy,Fz,Tx,Ty,Tz], N and Nm; measured parent-on-child reaction, child joint frame, common mount origin; no gravity tare.
- foot_state[B,10]: spherical foot center position relative to payload origin in rotating root frame(3,m); point velocity relative to translating AND rotating root in root frame(3,m/s; subtract omega cross r); ground normal contact force in WORLD frame(3,N); world foot-center z minus foot radius(1,m).
- guide_wrench[B,6]: actual passive guide force/torque in WORLD frame, mount origin, N/Nm; Fz exactly0.
- goal[B,1]: fixed nominal payload mounting-origin height in m, binding.height_m.
The loader subtracts goal from base_state z internally; do not pre-subtract it. Foot force is normal-only, not full friction wrench. Shapes alone cannot verify sensor frame correctness.
Binary file not shown.
@@ -0,0 +1,147 @@
{
"task_sha256": "bb2671ba1f1750baf6f0ec3ebb2f555d60ee252358867aa378fc85e07790c99a",
"support_sha256": "9910845796906fc79094fbe9d48600291bb90d4d4abb1b02d696747f8ecccae0",
"source_manifest_sha256": "4cea30b0264444c2e0417d0c26d6ea9148698d3b77f559f12a9b33775a41e3eb",
"training_identity": {
"config": "961bea90c341692a1b47f53eb0d4cff0746f4aed4a3a178d013c6dd8314d9a9e",
"sources": {
"scripts/train_loaded_stand.py": "40de1c7b540b6a7df096508d7d533b679b7d8352b84dfe25b4dc201cec4f5652",
"scripts/runtime_common.py": "57934860bc6fe1df64128376737aa2a80aafe8f85cf4bee38d7f4373edfe908b",
"scripts/supervise_sim.py": "1d917d29f87cde38cd75343285ec18fdebc19bc7329c89d782b172ffb4779134",
"source/SkillFactory/SkillFactory/legs/__init__.py": "ca6f7541ea78f762be54147f3c978ed9564d045115fdfc975639e639696b8876",
"source/SkillFactory/SkillFactory/legs/frozen_policy.py": "562d4ff6e4b5d1d569e16badcc44c972d303284e06371be22060a682aae38ffb",
"source/SkillFactory/SkillFactory/legs/spec.py": "e11d0df51131162c2416b099c2263e554ca72f13000a04828c79afc85a63f01d",
"source/SkillFactory/SkillFactory/legs/policy.py": "cdfcc1da937d1b33652601d52a76b7c253ef64179822133317b0918ea84c2f6a",
"source/SkillFactory/SkillFactory/legs/urdf.py": "694a6e0cf1c0825f02788982f528481932fe5cffb80738f130497c1040125c5a",
"source/SkillFactory/SkillFactory/legs/tokens.py": "86dccd0a4af945700ff3382d2638d9ca8ebe06151c403d90b4cdb00a8fde7ad5",
"source/SkillFactory/SkillFactory/legs/offline.py": "12ac018a2c76d046c8ed50ee105fcd6a33931f7c7a3bb027b6db332726ab064d",
"source/SkillFactory/SkillFactory/legs/standing_ppo.py": "69d99ed40369a07ec1c8fee4220d9adbeb38958499d94c1bd7da054b78598565",
"source/SkillFactory/SkillFactory/legs/support.py": "2abb489659001ba12e23becdbb02bec45f69a2e0b4d7efaeaec1c72e25d888cc",
"source/SkillFactory/SkillFactory/tasks/manager_based/loaded_leg/__init__.py": "cabfbd1a302f26e808bf61b2d672eae8f29fd31fbeb2c4e559d1e5dc2631d28c",
"source/SkillFactory/SkillFactory/tasks/manager_based/loaded_leg/cfg.py": "36987e4eaf8973925ec7025fd29f458de3cef7f667fe794a989bbba8817df70f",
"source/SkillFactory/SkillFactory/tasks/manager_based/loaded_leg/mdp.py": "83ec027cf39675749c25ed0fe01518f8c430a768a5465e4c68efdfd5da6884da",
"source/SkillFactory/SkillFactory/tasks/manager_based/loaded_leg/env.py": "f18f82fe0fa50ef5591a247431d9230440021635c7025f0c5f458ef266ded855",
"source/SkillFactory/SkillFactory/tasks/manager_based/leg_skill/__init__.py": "85417fb746517b22ec450803a02b09e23bf06421d0e78d83f272cff7de2eb678",
"source/SkillFactory/SkillFactory/tasks/manager_based/leg_skill/leg_env_cfg.py": "1c748307cc08e1b8367bb92fdf13a45bc3204abdf1cf423571b06d071e1b4815",
"source/SkillFactory/SkillFactory/tasks/manager_based/leg_skill/leg_env.py": "abf20f5feb2114c02b5fc8647102430881640a0d3f3ac2a6503ff0dd4395cd98",
"source/SkillFactory/SkillFactory/tasks/manager_based/leg_skill/mdp.py": "3ac1792858de1f7576556661e4cfb8eb55c348e7fa3cf61c8e62d7fe04969bc0",
"source/SkillFactory/SkillFactory/tasks/manager_based/single_leg_anchor/mdp/__init__.py": "2b998fd206546dc11d0ee693ab1030a3d9949094a6ee25d07bdabefab84a3806",
"source/SkillFactory/SkillFactory/tasks/manager_based/single_leg_anchor/mdp/observations.py": "019a971445b99708386a6ffb27ec837f3b0aef1aa8fea59e8ae638783769b1f3",
"source/SkillFactory/SkillFactory/tasks/manager_based/single_leg_anchor/mdp/actions.py": "91aa6bd581437c932c19415c7c1abc45b2329b9e72a10d17aebeda6142a249da",
"configs/standing/go2_fl_support_v0.json": "bb2671ba1f1750baf6f0ec3ebb2f555d60ee252358867aa378fc85e07790c99a",
"source/SkillFactory/SkillFactory/legs/resources/ground.usda": "26e7ef57327a36b59aacd0cc1ba0704ec4d86387450b4cab1f4020ae28bac70d",
"assets/generated/support/go2_fl_3kg_v0/support.json": "9910845796906fc79094fbe9d48600291bb90d4d4abb1b02d696747f8ecccae0",
"assets/generated/legs/go2_fl/manifest.json": "4cea30b0264444c2e0417d0c26d6ea9148698d3b77f559f12a9b33775a41e3eb"
},
"features": "loaded-stand-45-fixed-si-v1",
"probe_episode_s": 0.0,
"device": "cuda:0",
"torch_version": "2.10.0+cu128"
},
"checkpoint_sha256": "f2197339dea3622ed92b82fdddff3400f2a82f3ca4d4fd3be6c72b45f1a77704",
"height_m": 0.3345983503570176,
"nominal": [
0.0,
0.8,
-1.5
],
"action_scale_rad": 0.5,
"soft_limits": [
[
-0.94248,
0.94248
],
[
-1.3177249999999998,
3.237625
],
[
-2.6284530000000004,
-0.9320069999999999
]
],
"joint_order": [
"FL_hip_joint",
"FL_thigh_joint",
"FL_calf_joint"
],
"control_dt": 0.01,
"physics_dt": 0.002,
"quaternion_order": "xyzw",
"solver": "PGS",
"mount_wrench": "parent-on-child; child joint frame at mount origin; untared N/Nm",
"limits": {
"scope": "guided single Go2 FL; synthetic3kg payload; zero vertical guide",
"not_claimed": [
"unsupported balance",
"full quadruped",
"Go1",
"hardware safety",
"generic v2 adapter"
]
},
"support_manifest": {
"contract": "go2-fl-loaded-stand-v0",
"fixed_to_world": false,
"mount_joint": "mount_ft_joint",
"source_manifest_sha256": "4cea30b0264444c2e0417d0c26d6ea9148698d3b77f559f12a9b33775a41e3eb",
"usd_path": "usd/loaded_leg/loaded_leg.usda",
"body_paths": {
"virtual_payload": "/Geometry/virtual_payload",
"anchor_base": "/Geometry/virtual_payload/anchor_base",
"FL_hip": "/Geometry/virtual_payload/anchor_base/FL_hip",
"FL_thigh": "/Geometry/virtual_payload/anchor_base/FL_hip/FL_thigh",
"FL_calf": "/Geometry/virtual_payload/anchor_base/FL_hip/FL_thigh/FL_calf",
"FL_calflower": "/Geometry/virtual_payload/anchor_base/FL_hip/FL_thigh/FL_calf/FL_calflower",
"FL_calflower1": "/Geometry/virtual_payload/anchor_base/FL_hip/FL_thigh/FL_calf/FL_calflower/FL_calflower1",
"FL_foot": "/Geometry/virtual_payload/anchor_base/FL_hip/FL_thigh/FL_calf/FL_foot",
"FL_calf_rotor": "/Geometry/virtual_payload/anchor_base/FL_hip/FL_thigh/FL_calf_rotor",
"FL_thigh_rotor": "/Geometry/virtual_payload/anchor_base/FL_hip/FL_thigh_rotor",
"FL_hip_rotor": "/Geometry/virtual_payload/anchor_base/FL_hip_rotor"
},
"joint_paths": {
"mount_ft_joint": "/Physics/mount_ft_joint",
"FL_hip_joint": "/Physics/FL_hip_joint",
"FL_hip_rotor_joint": "/Physics/FL_hip_rotor_joint",
"FL_thigh_joint": "/Physics/FL_thigh_joint",
"FL_thigh_rotor_joint": "/Physics/FL_thigh_rotor_joint",
"FL_calf_joint": "/Physics/FL_calf_joint",
"FL_calf_rotor_joint": "/Physics/FL_calf_rotor_joint",
"FL_calflower_joint": "/Physics/FL_calflower_joint",
"FL_foot_joint": "/Physics/FL_foot_joint",
"FL_calflower1_joint": "/Physics/FL_calflower1_joint"
},
"missing_source_inertial_links": [
"FL_calflower",
"FL_calflower1"
],
"payload_mass_kg": 3.0,
"payload_size_m": [
0.18,
0.18,
0.08
],
"payload_com_m": [
0.0,
0.0,
0.06
],
"sensor_carrier_mass_kg": 0.05,
"files": {
"LICENSE": "84aac59fd3246e3ddc49d1387644e8fcf43b0def4f0b9fe687f372e90446df2d",
"dae/calf.dae": "29872fa1a435b92bc2b395fdfa7f0aee26101db3c494681896bd4ef24ee7883c",
"dae/foot.dae": "ffbd95fd641866ce9bd277aa249b92dafc9148d70131fa2c52709a35ce5b710b",
"dae/hip.dae": "8735e10617afe252cef0844c1d139d15bc6d4c653660d742fea9fe05c2704d8f",
"dae/thigh.dae": "a622381ae308897ff143b8bbf02f075457c342b22984205040424a5a983246fa",
"loaded_leg.urdf": "d9ed98ca72012159b027bafacf6181bb4276ed849d270081166e006973bdf56b",
"usd/loaded_leg/loaded_leg.usda": "fe888c68398bf41e8644ba9b894a153042807e4c5c881a8e3bc3ed9f2a2b84ff",
"usd/loaded_leg/payloads/Physics/physics.usda": "091d14a16869bc833369835ea643e0cd231e98dab988761b47904bf0c47da56a",
"usd/loaded_leg/payloads/Physics/physx.usda": "bc384acc4cd3ffefd0202c11fdea0a4448adb4c912ebbf52197a96d304dbe42e",
"usd/loaded_leg/payloads/base.usda": "ef7b19d2bcd76d60b894ab368f141b80beaf9c0c69f37596a6fc6b774f7dc4f8",
"usd/loaded_leg/payloads/geometries.usd": "68771be92510be58f2b0ce8891089d08a5eaee7a80352fca4e84808b7ebd0034",
"usd/loaded_leg/payloads/instances.usda": "eac5ff19064fd0aedb49d6975936aa4f0072438fb6e6593084a0bcf39581e403",
"usd/loaded_leg/payloads/materials.usda": "3bf7b2e12dde571af09be43b36a7b3825e96f0578d0b97dc85118efdd174115c",
"usd/loaded_leg/payloads/robot.usda": "187c18cc30e67098b14ac144aa6b4ec0c3e960febb8dc7f29419cbd87b469f3f"
}
}
}
@@ -0,0 +1,292 @@
{
"trials": {
"full": {
"passed": true,
"accepted_by_operator_waiver": false,
"gates": {
"finite_complete_trace": true,
"floating_root": true,
"pgs_compatible": true,
"no_autoreset": true,
"no_fall": true,
"applied_load_matches_protocol": true,
"quaternion_norm": true,
"no_vertical_guide": true,
"height": true,
"tilt": true,
"foot_drift": true,
"contact": true,
"vertical_balance": true,
"limits": true,
"sensor_static_force": true,
"sensor_static_torque": true,
"indexed_reset": true,
"timeout_terminal_reset": true,
"bad_action_atomic_rejection": true
},
"mean_height_error_m": 0.0010973225607199789,
"cpu_cuda_max_error": 1.7881393432617188e-07,
"physical_samples": 7000,
"max_height_error_m": 0.011397260259590258,
"max_tilt_rad": 0.039559598305326425,
"max_foot_drift_m": 0.004637469188639254,
"min_contact_fraction": 1.0,
"sensor_force_error_N": 0.029688779830932432,
"sensor_torque_error_Nm": 0.001858974476667002,
"memory_growth_fraction": null
},
"holdout": {
"passed": true,
"accepted_by_operator_waiver": false,
"gates": {
"finite_complete_trace": true,
"floating_root": true,
"pgs_compatible": true,
"no_autoreset": true,
"no_fall": true,
"applied_load_matches_protocol": true,
"quaternion_norm": true,
"no_vertical_guide": true,
"height": true,
"tilt": true,
"foot_drift": true,
"contact": true,
"vertical_balance": true,
"limits": true
},
"mean_height_error_m": 0.0009209314090133883,
"cpu_cuda_max_error": 1.7881393432617188e-07,
"physical_samples": 7000,
"max_height_error_m": 0.01069288236995708,
"max_tilt_rad": 0.03849515058506419,
"max_foot_drift_m": 0.004441969886462753,
"min_contact_fraction": 1.0,
"sensor_force_error_N": 0.11419118010997775,
"sensor_torque_error_Nm": 0.007193774832685795,
"memory_growth_fraction": null
},
"hold": {
"passed": false,
"accepted_by_operator_waiver": false,
"gates": {
"finite_complete_trace": true,
"floating_root": true,
"pgs_compatible": true,
"no_autoreset": true,
"no_fall": true,
"applied_load_matches_protocol": true,
"quaternion_norm": true,
"no_vertical_guide": true,
"height": false,
"tilt": true,
"foot_drift": true,
"contact": true,
"vertical_balance": true,
"limits": true
},
"mean_height_error_m": 0.06637654078181726,
"cpu_cuda_max_error": 0.0,
"physical_samples": 7000,
"max_height_error_m": 0.09780764861007701,
"max_tilt_rad": 0.05186780450752031,
"max_foot_drift_m": 0.0033128396003425657,
"min_contact_fraction": 1.0,
"sensor_force_error_N": 0.12399852900206909,
"sensor_torque_error_Nm": 0.005068768669592938,
"memory_growth_fraction": null
},
"stability": {
"passed": false,
"accepted_by_operator_waiver": true,
"gates": {
"finite_complete_trace": true,
"floating_root": true,
"pgs_compatible": true,
"no_autoreset": true,
"no_fall": true,
"applied_load_matches_protocol": true,
"quaternion_norm": true,
"no_vertical_guide": true,
"height": true,
"tilt": true,
"foot_drift": false,
"contact": true,
"vertical_balance": true,
"limits": true,
"sensor_static_force": true,
"sensor_static_torque": true,
"memory": true
},
"mean_height_error_m": 0.0017030573373548721,
"cpu_cuda_max_error": 2.384185791015625e-07,
"physical_samples": 60000,
"max_height_error_m": 0.020278918575248828,
"max_tilt_rad": 0.057941022266468155,
"max_foot_drift_m": 0.03078258298377699,
"min_contact_fraction": 1.0,
"sensor_force_error_N": 0.02974697828292827,
"sensor_torque_error_Nm": 0.0018565384255236957,
"memory_growth_fraction": 0.0
},
"exploratory": {
"passed": true,
"accepted_by_operator_waiver": false,
"gates": {
"finite_complete_trace": true,
"floating_root": true,
"pgs_compatible": true,
"no_autoreset": true,
"no_fall": true,
"applied_load_matches_protocol": true,
"quaternion_norm": true,
"no_vertical_guide": true,
"height": true,
"tilt": true,
"foot_drift": true,
"contact": true,
"vertical_balance": true,
"limits": true
},
"mean_height_error_m": 0.001673169419981864,
"cpu_cuda_max_error": 1.7881393432617188e-07,
"physical_samples": 7000,
"max_height_error_m": 0.019595432113609423,
"max_tilt_rad": 0.05547213523615746,
"max_foot_drift_m": 0.007914748932079697,
"min_contact_fraction": 1.0,
"sensor_force_error_N": 0.05928855037689118,
"sensor_torque_error_Nm": 0.003488774794082272,
"memory_growth_fraction": null
}
},
"holdout_improvement_fraction": 0.9861256492404368,
"artifacts": {
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/full/result.json": "449f62b55ada8c84c6d67783c2891684984b1a6ed5ec5aff69aa08ee14b083df",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/full/metadata.json": "ba05777054e2f87a0be328392b5beccf3e95f03e5431855ba0f4b6b8142c82c7",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/full/physics_trace.pt": "5759859ad3b8b41d55c906925c8812101b6511843785350acba08fe48c3217cf",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/full/inference_samples.pt": "106a23047f252397f3e853e661aff99565bcb0867a47da40227659ccfd1224f0",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/full/memory.json": "37517e5f3dc66819f61f5a7bb8ace1921282415f10551d2defa5c3eb0985b570",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/full_process.json": "26712c9dc6c0e17bc5614ad2b7d39906a3e111c2d00d413147732089d4308767",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/full.log": "3ed37ffa54cf29db698adb0a4fb117494982c821820efc34767dbb2d9be46c08",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/holdout/result.json": "725ed47c750053f594236b1f611550a3e234fa82b546a84da3831266b589b8d5",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/holdout/metadata.json": "3a8fa74aa6178fde5c646c993706736fcc7250cf5b7d097dd8f223571b72960b",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/holdout/physics_trace.pt": "c30e9e1ced212947ab7ec553020c64793e69d830488ecc242881efa7e8fb509f",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/holdout/inference_samples.pt": "c93ba31e76755a443dba518ffd97bebe520e67bdf8f0739004158227e35bdf14",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/holdout/memory.json": "37517e5f3dc66819f61f5a7bb8ace1921282415f10551d2defa5c3eb0985b570",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/holdout_process.json": "0750f7d89d7058b8873f54909f32429a0977dfd65c4e3b9623c93b620c79c58b",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/holdout.log": "afdea6d441b587e04527010b0989e9541e063cc6bc035b9d8b6ecac80413f248",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/hold/result.json": "aeb05e1fb283d5bbcad54ceec968031907b60d5c424fc585c8da71d50d844d1a",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/hold/metadata.json": "3a8fa74aa6178fde5c646c993706736fcc7250cf5b7d097dd8f223571b72960b",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/hold/physics_trace.pt": "6bdd52a425fe1a92d5726ce12e6ce460902cbbf121160d782bd6d027b8b4d91c",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/hold/inference_samples.pt": "33e4805548666672b8615a82034f0638ed53bc9af6ecfeabd6aa0e16d7231d5b",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/hold/memory.json": "37517e5f3dc66819f61f5a7bb8ace1921282415f10551d2defa5c3eb0985b570",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/hold_process.json": "3b08d09a04bcdef07ad198cf9654b74edbb0d54eea6043031d52e7d852f04e78",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/hold.log": "17648aed5613a3516e34e32d6d53ddfb6f832e66e2745f035e928bf91a007e3b",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/stability/result.json": "95a0f18edac40ece69430041b5083a011fe612a4770e899575ca985ba496e0b4",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/stability/metadata.json": "4c26c6fc3f2eb59b0fe6f2a5d1962c283f2da8e2c4047a33929f6643b17ab447",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/stability/physics_trace.pt": "3b35f32f71e87c812a24ae93101f05c917707691fea15ee523cca161846fe498",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/stability/inference_samples.pt": "3b21263ef59eef914c2bd5e38a86f98145593a9098fa7eb014c14abff6befc17",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/stability/memory.json": "3360e264f6062eff14fd1a015e670c26bdc1f84d1b7acc9dd568c35fd63868c5",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/stability_process.json": "e2af5a4fd4341dcb4ccecdef54dfa9c6404133a54cf2d7d05179056095ddd26d",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/stability.log": "e7678de92de394288d0fa5e570de090cb323bbd23411a49a7390f6ef30c6f380",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/exploratory/result.json": "b7d6aac44a871ce33c0bfa6b5c8578749d62b882a516c54bda36afd3d3cce5e8",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/exploratory/metadata.json": "267e15afe4e32636cefcedb047b8acc3958f607a7ea234b85574c5f919e2f7d2",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/exploratory/physics_trace.pt": "eb90f662c1e485bf993df4d4d3b3d8e1c885dd8537c78cd3222f4f2c217e9726",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/exploratory/inference_samples.pt": "f3d9cf3d7bd95bc938a19ae3bd27618da597e62d1168d2a64543e73300d2aec0",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/exploratory/memory.json": "37517e5f3dc66819f61f5a7bb8ace1921282415f10551d2defa5c3eb0985b570",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/exploratory_process.json": "350ec331451ddb9b41d3386fe4414214242cd799d8ae72c881a296ccf23f6e70",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/exploratory.log": "251570e2b828b434d7e1238fd4a826f72fc9435a5bbb6ebb127ba39cd727d251",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/release_deviation_v1/operator_waiver.json": "3c8c03efcedc3c6eb2a0a8ca8fdae2629834473fb122b2ce5bbcc39161344d1e",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/release_deviation_v1/engineering.json": "430ab38191962e1da218d4375c4a489019ac0d5e3e04333d236c41076a2b7cef",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/candidate_lock.json": "1dce2b41208f438bd98958d443dd8106574eb19063e66b9cfddf9edfbec258b4",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/independent_candidate.json": "54396bb01be4c499b41a328547614980cf702232fd57c014ec6c37d96561fd3a"
},
"candidate_manifest_sha256": "7cb478a3b17aaf307d2d6e3c2b49cc25aa9af1ee9d118dd731b52f618c27e628",
"core_files": {
"ASSET_LICENSE": "84aac59fd3246e3ddc49d1387644e8fcf43b0def4f0b9fe687f372e90446df2d",
"README.md": "2c8cb0ea4f057b5c07600f2054aa58eed35b5e45e253ce7eec9bc1a532fc57ea",
"actor.pt": "1384b5aa780cded1f4ed21b617ff8c883a5f82826c4cb4c9415b75f315b1e664",
"binding.json": "22cd42a44f29913b2f0bd04e685e77acdcc1ef7b0218f62733dc401d55b6d710",
"example.pt": "f21e71bbc4b0e862a672d13daf67c31971f365426dd98662fc658ede6916c798",
"frozen_standing.py": "c19026845ea5ec98b21ed641cd294cccd502cf73907c08f0fe71a9a63067d8c6",
"requirements.txt": "812ea7d306f227902500eb76d816fe9f65f75d201edc6e766118ed68d3fdfe0d",
"selection.json": "276302e9374337aed8d4d3ea35408c4a0dc9971ea37bcecda8ccc5208b6a296a",
"task.json": "bb2671ba1f1750baf6f0ec3ebb2f555d60ee252358867aa378fc85e07790c99a",
"training.json": "443187246b6a600a11566eb1f1d3dca260f7c6a64684a847af7577afc4e521c9"
},
"acceptance_basis": "operator_waiver",
"unwaived_required_physics_passed": false,
"waiver": {
"format": "skillfactory.standing-deviation.v1",
"authorization": "explicit_user_instruction",
"operator_statement": "\u8fd9\u4e2a120\u79d2\u8010\u4e45\u7684\u98d8\u9038\u8d85\u8fc7\u7684\u5f88\u5c0f\uff0c\u53ef\u4ee5\u89c6\u4e3a\u901a\u8fc7",
"recorded_at": "2026-10-12T16:58:57.445471+08:00",
"approved": true,
"scope": "exact_candidate_and_trial_only",
"waived_gate": "stability.foot_drift",
"original_limit_m": 0.03,
"candidate_manifest_sha256": "7cb478a3b17aaf307d2d6e3c2b49cc25aa9af1ee9d118dd731b52f618c27e628",
"actor_sha256": "1384b5aa780cded1f4ed21b617ff8c883a5f82826c4cb4c9415b75f315b1e664",
"checkpoint_sha256": "f2197339dea3622ed92b82fdddff3400f2a82f3ca4d4fd3be6c72b45f1a77704",
"duration_s": 120.0,
"num_envs": 26,
"physical_samples": 60000,
"observed_failures": [
{
"environment": 6,
"max_foot_drift_m": 0.030620924223339147
},
{
"environment": 19,
"max_foot_drift_m": 0.03078258298377699
}
],
"evidence_sha256": {
"stability/result.json": "95a0f18edac40ece69430041b5083a011fe612a4770e899575ca985ba496e0b4",
"stability/metadata.json": "4c26c6fc3f2eb59b0fe6f2a5d1962c283f2da8e2c4047a33929f6643b17ab447",
"stability/physics_trace.pt": "3b35f32f71e87c812a24ae93101f05c917707691fea15ee523cca161846fe498",
"stability/memory.json": "3360e264f6062eff14fd1a015e670c26bdc1f84d1b7acc9dd568c35fd63868c5",
"stability_process.json": "e2af5a4fd4341dcb4ccecdef54dfa9c6404133a54cf2d7d05179056095ddd26d",
"stability.log": "e7678de92de394288d0fa5e570de090cb323bbd23411a49a7390f6ef30c6f380"
},
"not_authorized": [
"changing_original_results",
"changing_global_threshold",
"waiving_other_gates",
"new_weights_or_trials",
"hardware_safety_claims"
]
},
"waiver_sha256": "3c8c03efcedc3c6eb2a0a8ca8fdae2629834473fb122b2ce5bbcc39161344d1e",
"waiver_receipts": [
{
"accepted_by_operator_waiver": true,
"raw_gate_passed": false,
"gate": "stability.foot_drift",
"original_limit_m": 0.03,
"observed_failures": [
{
"environment": 6,
"max_foot_drift_m": 0.030620924223339147
},
{
"environment": 19,
"max_foot_drift_m": 0.03078258298377699
}
]
}
],
"passed": true,
"status": "completed",
"accepted_skill": false,
"stage": "eligible_for_frozen_publication",
"training": {
"updates": 2048,
"transitions": 16777216,
"selected_update": 1024,
"checkpoint_sha256": "f2197339dea3622ed92b82fdddff3400f2a82f3ca4d4fd3be6c72b45f1a77704"
},
"physics_wall_seconds": 502.72002290599266,
"artifact_bytes": 1721418294,
"engineering_path": "/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/release_deviation_v1/engineering.json"
}
@@ -0,0 +1,20 @@
{
"passed": true,
"accepted_skill": false,
"checkpoint_sha256": "f2197339dea3622ed92b82fdddff3400f2a82f3ca4d4fd3be6c72b45f1a77704",
"selected_update": 1024,
"manifest_sha256": "7cb478a3b17aaf307d2d6e3c2b49cc25aa9af1ee9d118dd731b52f618c27e628",
"core_files": {
"ASSET_LICENSE": "84aac59fd3246e3ddc49d1387644e8fcf43b0def4f0b9fe687f372e90446df2d",
"README.md": "2c8cb0ea4f057b5c07600f2054aa58eed35b5e45e253ce7eec9bc1a532fc57ea",
"actor.pt": "1384b5aa780cded1f4ed21b617ff8c883a5f82826c4cb4c9415b75f315b1e664",
"binding.json": "22cd42a44f29913b2f0bd04e685e77acdcc1ef7b0218f62733dc401d55b6d710",
"example.pt": "f21e71bbc4b0e862a672d13daf67c31971f365426dd98662fc658ede6916c798",
"frozen_standing.py": "c19026845ea5ec98b21ed641cd294cccd502cf73907c08f0fe71a9a63067d8c6",
"requirements.txt": "812ea7d306f227902500eb76d816fe9f65f75d201edc6e766118ed68d3fdfe0d",
"selection.json": "276302e9374337aed8d4d3ea35408c4a0dc9971ea37bcecda8ccc5208b6a296a",
"task.json": "bb2671ba1f1750baf6f0ec3ebb2f555d60ee252358867aa378fc85e07790c99a",
"training.json": "443187246b6a600a11566eb1f1d3dca260f7c6a64684a847af7577afc4e521c9"
},
"actor_extraction_max_error": 0.0
}
@@ -0,0 +1,26 @@
{
"passed": true,
"scope": "engineering and explicit deviation handling only",
"accepted_skill": false,
"review_passed": true,
"pytest_log": "logs/loaded_stand/release_deviation_v1/pytest_final.log",
"ruff_log": "logs/loaded_stand/release_deviation_v1/ruff.log",
"review_report": "logs/loaded_stand/release_deviation_v1/review.md",
"files": {
"scripts/standing_acceptance.py": "a5cfbde5fe174973e1081d02b35ff3d8230ed77f963c729f24c43a43e78750c9",
"scripts/standing_deviation.py": "33f50f25e3099297cca00ccf120b66c09fd9c43ba88c774219f53a0028946467",
"scripts/publish_standing.py": "7fbdfc6e5e0c226beb87d513f18f448683b042f48397721b66a17b7ed0368025",
"scripts/verify_frozen_standing.py": "7528979d2281ed3831f2cd3fadaba55a62a5a927ec33c587e43832dded51c409",
"scripts/package_standing.py": "79d225924caf78b71ccfc348042977046bcd36b036933d0091ba09a5a95ccdb7",
"scripts/evaluate_standing.py": "692ac51e07969565381ff49d08c82524bf275bfc50d555543fb20420992b96ad",
"scripts/standing_protocol.py": "d8ba91c96006687ba5d3e54fd986c915a034f16f11748ea06d717632b6322db1",
"source/SkillFactory/SkillFactory/legs/frozen_standing.py": "c19026845ea5ec98b21ed641cd294cccd502cf73907c08f0fe71a9a63067d8c6",
"tests/test_standing_acceptance.py": "2dc4339582c911c51c9f1cb3dbcc50e6581a023d6d3dedea764752d7d206f47f",
"tests/test_standing_deviation.py": "33bbf43a4dd0ed10428ac156c35033ab52aecc448d7a35edf7e2cbbc4e5ee9c6",
"logs/loaded_stand/release_deviation_v1/pytest_final.log": "911005ae3371eefae02b59c53e82d3fb9ce75ab4e873ee8d51ecaef3e0d93eff",
"logs/loaded_stand/release_deviation_v1/ruff.log": "a5b097bbb606a1132331f167e0dab21849bbe45b37cc1957f17dcc665ece6218",
"logs/loaded_stand/release_deviation_v1/review.md": "e04acdcc57bf3365fd05d5b224f384f94d2f55d60391af16efded1edaa932d41",
"logs/loaded_stand/release_deviation_v1/operator_waiver.json": "3c8c03efcedc3c6eb2a0a8ca8fdae2629834473fb122b2ce5bbcc39161344d1e"
},
"pytest": "256 passed, 3 skipped"
}
File diff suppressed because it is too large Load Diff
@@ -0,0 +1,40 @@
{
"passed": true,
"status": "completed",
"command": [
"/home/cen/IsaacLab/isaaclab.sh",
"-p",
"scripts/evaluate_standing.py",
"--package",
"logs/loaded_stand/acceptance_v1/v2/candidate",
"--manifest",
"7cb478a3b17aaf307d2d6e3c2b49cc25aa9af1ee9d118dd731b52f618c27e628",
"--mode",
"exploratory",
"--output",
"logs/loaded_stand/acceptance_v1/v2/exploratory",
"--headless"
],
"pid": 350390,
"cleanup": {
"before": [
{
"pid": 350475,
"state": "Z",
"pgid": 350390,
"name": "Omniverse Hub"
},
{
"pid": 350485,
"state": "S",
"pgid": 350485,
"name": "omni.telemetry."
}
],
"remaining": [],
"clean": true
},
"returncode": 0,
"wall_seconds": 38.74840120600129,
"critical_log_lines": []
}
@@ -0,0 +1,448 @@
{
"passed": true,
"gates": {
"finite_complete_trace": true,
"floating_root": true,
"pgs_compatible": true,
"no_autoreset": true,
"no_fall": true,
"applied_load_matches_protocol": true,
"quaternion_norm": true,
"no_vertical_guide": true,
"height": true,
"tilt": true,
"foot_drift": true,
"contact": true,
"vertical_balance": true,
"limits": true,
"sensor_static_force": true,
"sensor_static_torque": true,
"indexed_reset": true,
"timeout_terminal_reset": true,
"bad_action_atomic_rejection": true
},
"physical_samples": 7000,
"mean_height_error_m": 0.0010973225607199789,
"rows": [
{
"mean_height_error_m": 0.0005724067638857842,
"max_height_error_m": 0.0006537912597275874,
"max_height_error_all_m": 0.004208910297355761,
"max_tilt_rad": 0.013293655852925941,
"max_foot_drift_m": 0.0003743763895022189,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 8.67843916037001e-06,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559163792118146,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.00150410074764714,
"max_height_error_m": 0.003285097907981982,
"max_height_error_all_m": 0.004209536146125903,
"max_tilt_rad": 0.019992917052446373,
"max_foot_drift_m": 0.0023749326706273906,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 8.913903018805542e-06,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559163792118146,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0006537433480539311,
"max_height_error_m": 0.0012591658343888423,
"max_height_error_all_m": 0.004209267925224414,
"max_tilt_rad": 0.015385177738147392,
"max_foot_drift_m": 0.00300550507152529,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 9.135441077998805e-06,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559163792118146,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0005594190504737396,
"max_height_error_m": 0.0014445660819626949,
"max_height_error_all_m": 0.0042097745647050044,
"max_tilt_rad": 0.01704234546342108,
"max_foot_drift_m": 0.0020779705784886317,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 8.572732964504666e-06,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559163792118146,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0019554498880327317,
"max_height_error_m": 0.004563230108223071,
"max_height_error_all_m": 0.004563230108223071,
"max_tilt_rad": 0.031200907071322788,
"max_foot_drift_m": 0.0016207407426390582,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 9.809723945019197e-06,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559165006989887,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0010418824156519437,
"max_height_error_m": 0.002552938629188428,
"max_height_error_all_m": 0.004210370611152758,
"max_tilt_rad": 0.017048920564464116,
"max_foot_drift_m": 0.0030618251865627156,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 9.740130033042778e-06,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559165006989887,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.00441482886532821,
"max_height_error_m": 0.011397260259590258,
"max_height_error_all_m": 0.011397260259590258,
"max_tilt_rad": 0.039559598305326425,
"max_foot_drift_m": 0.004637469188639254,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 8.302421528345711e-06,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559164399554016,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0005586289034245191,
"max_height_error_m": 0.0006710170020676753,
"max_height_error_all_m": 0.0042078076114274165,
"max_tilt_rad": 0.022379360046486673,
"max_foot_drift_m": 0.0003796737131488018,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 9.035651542821874e-06,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559165006989887,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0006638780769850724,
"max_height_error_m": 0.0008434830417252681,
"max_height_error_all_m": 0.004210638832054248,
"max_tilt_rad": 0.013611028098504098,
"max_foot_drift_m": 0.0005755334252334962,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 8.909297162961809e-06,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559165006989887,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0008461116204652958,
"max_height_error_m": 0.0012549339046097896,
"max_height_error_all_m": 0.004209267925224414,
"max_tilt_rad": 0.020649614222536222,
"max_foot_drift_m": 0.00040588496126681476,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 8.91819620557539e-06,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559165006989887,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.00034732328476711697,
"max_height_error_m": 0.0006549535503007076,
"max_height_error_all_m": 0.004209536146125903,
"max_tilt_rad": 0.013297898156648192,
"max_foot_drift_m": 0.0003790252940569456,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 9.224287195724217e-06,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559165006989887,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0006014416503993097,
"max_height_error_m": 0.0006598113288499019,
"max_height_error_all_m": 0.004212397169075122,
"max_tilt_rad": 0.013299008300830244,
"max_foot_drift_m": 0.0003801550413455481,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 9.139830470554821e-06,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559165006989887,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0005459786742449322,
"max_height_error_m": 0.0006531654109574458,
"max_height_error_all_m": 0.004209804367027392,
"max_tilt_rad": 0.01329671728403836,
"max_foot_drift_m": 0.00037841122110757435,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 9.05201856597165e-06,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559163792118146,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
}
],
"load_reconstruction_max_error": 0.0,
"calibration_counts": [
500,
500,
500
],
"sensor_static_max_force_error_N": 0.029688779830932432,
"sensor_static_max_torque_error_Nm": 0.001858974476667002,
"mode": "full",
"controller": "policy",
"input_hash": "3105c4fbb376912d5d397055995e94070eeb6a3e59f4c5fae2e416f6e84a0da9",
"metadata": {
"inputs": {
"mode": "full",
"count": 13,
"duration_s": 14.0,
"seed": 71000,
"basis": [
[
0.0,
0.0,
0.0,
0.0,
0.0,
0.0
],
[
1.0,
0.0,
0.0,
0.0,
0.0,
0.0
],
[
-1.0,
0.0,
0.0,
0.0,
0.0,
0.0
],
[
0.0,
1.0,
0.0,
0.0,
0.0,
0.0
],
[
0.0,
-1.0,
0.0,
0.0,
0.0,
0.0
],
[
0.0,
0.0,
1.0,
0.0,
0.0,
0.0
],
[
0.0,
0.0,
-1.0,
0.0,
0.0,
0.0
],
[
0.0,
0.0,
0.0,
1.0,
0.0,
0.0
],
[
0.0,
0.0,
0.0,
-1.0,
0.0,
0.0
],
[
0.0,
0.0,
0.0,
0.0,
1.0,
0.0
],
[
0.0,
0.0,
0.0,
0.0,
-1.0,
0.0
],
[
0.0,
0.0,
0.0,
0.0,
0.0,
1.0
],
[
0.0,
0.0,
0.0,
0.0,
0.0,
-1.0
]
],
"scale": [
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0
],
"shift_s": [
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0
],
"period_s": 14.0
},
"input_hash": "3105c4fbb376912d5d397055995e94070eeb6a3e59f4c5fae2e416f6e84a0da9",
"physics_dt": 0.002,
"control_dt": 0.01,
"height_m": 0.3345983503570176,
"floating_root": true,
"solver": "PGS",
"tgs_force_flag": false,
"autoresets": 0,
"total_mass_kg": 5.388350963592529,
"mass_kg": {
"virtual_payload": 3.0,
"anchor_base": 0.05000000074505806,
"FL_hip": 0.6779999732971191,
"FL_hip_rotor": 0.08900000154972076,
"FL_thigh": 1.1519999504089355,
"FL_thigh_rotor": 0.08900000154972076,
"FL_calf": 0.15399999916553497,
"FL_calf_rotor": 0.08900000154972076,
"FL_calflower": 0.02470862679183483,
"FL_foot": 0.03999999910593033,
"FL_calflower1": 0.02264302782714367
},
"limits": [
[
-1.0472,
1.0472
],
[
-1.5708,
3.4907
],
[
-2.7227,
-0.83776
]
],
"speeds": [
30.1,
30.1,
15.7
],
"efforts": [
23.700000762939453,
23.700000762939453,
45.43000030517578
]
},
"accepted_skill": false,
"package_manifest_sha256": "7cb478a3b17aaf307d2d6e3c2b49cc25aa9af1ee9d118dd731b52f618c27e628",
"actor_sha256": "1384b5aa780cded1f4ed21b617ff8c883a5f82826c4cb4c9415b75f315b1e664",
"physics_trace_sha256": "5759859ad3b8b41d55c906925c8812101b6511843785350acba08fe48c3217cf",
"source_hashes": {
"scripts/runtime_common.py": "57934860bc6fe1df64128376737aa2a80aafe8f85cf4bee38d7f4373edfe908b",
"source/SkillFactory/SkillFactory/tasks/manager_based/single_leg_anchor/mdp/__init__.py": "2b998fd206546dc11d0ee693ab1030a3d9949094a6ee25d07bdabefab84a3806",
"source/SkillFactory/SkillFactory/legs/resources/ground.usda": "26e7ef57327a36b59aacd0cc1ba0704ec4d86387450b4cab1f4020ae28bac70d",
"scripts/standing_protocol.py": "d8ba91c96006687ba5d3e54fd986c915a034f16f11748ea06d717632b6322db1",
"scripts/supervise_sim.py": "1d917d29f87cde38cd75343285ec18fdebc19bc7329c89d782b172ffb4779134",
"scripts/evaluate_standing.py": "692ac51e07969565381ff49d08c82524bf275bfc50d555543fb20420992b96ad",
"assets/generated/support/go2_fl_3kg_v0/support.json": "9910845796906fc79094fbe9d48600291bb90d4d4abb1b02d696747f8ecccae0",
"assets/generated/legs/go2_fl/manifest.json": "4cea30b0264444c2e0417d0c26d6ea9148698d3b77f559f12a9b33775a41e3eb",
"source/SkillFactory/SkillFactory/tasks/manager_based/loaded_leg/cfg.py": "36987e4eaf8973925ec7025fd29f458de3cef7f667fe794a989bbba8817df70f",
"source/SkillFactory/SkillFactory/legs/standing_ppo.py": "69d99ed40369a07ec1c8fee4220d9adbeb38958499d94c1bd7da054b78598565",
"source/SkillFactory/SkillFactory/legs/spec.py": "e11d0df51131162c2416b099c2263e554ca72f13000a04828c79afc85a63f01d",
"source/SkillFactory/SkillFactory/legs/__init__.py": "ca6f7541ea78f762be54147f3c978ed9564d045115fdfc975639e639696b8876",
"source/SkillFactory/SkillFactory/legs/frozen_policy.py": "562d4ff6e4b5d1d569e16badcc44c972d303284e06371be22060a682aae38ffb",
"source/SkillFactory/SkillFactory/tasks/manager_based/leg_skill/leg_env_cfg.py": "1c748307cc08e1b8367bb92fdf13a45bc3204abdf1cf423571b06d071e1b4815",
"source/SkillFactory/SkillFactory/tasks/manager_based/leg_skill/leg_env.py": "abf20f5feb2114c02b5fc8647102430881640a0d3f3ac2a6503ff0dd4395cd98",
"configs/standing/go2_fl_support_v0.json": "bb2671ba1f1750baf6f0ec3ebb2f555d60ee252358867aa378fc85e07790c99a",
"source/SkillFactory/SkillFactory/tasks/manager_based/single_leg_anchor/mdp/actions.py": "91aa6bd581437c932c19415c7c1abc45b2329b9e72a10d17aebeda6142a249da",
"source/SkillFactory/SkillFactory/tasks/manager_based/single_leg_anchor/mdp/observations.py": "019a971445b99708386a6ffb27ec837f3b0aef1aa8fea59e8ae638783769b1f3",
"scripts/train_loaded_stand.py": "40de1c7b540b6a7df096508d7d533b679b7d8352b84dfe25b4dc201cec4f5652",
"source/SkillFactory/SkillFactory/legs/policy.py": "cdfcc1da937d1b33652601d52a76b7c253ef64179822133317b0918ea84c2f6a",
"source/SkillFactory/SkillFactory/legs/offline.py": "12ac018a2c76d046c8ed50ee105fcd6a33931f7c7a3bb027b6db332726ab064d",
"source/SkillFactory/SkillFactory/tasks/manager_based/loaded_leg/mdp.py": "83ec027cf39675749c25ed0fe01518f8c430a768a5465e4c68efdfd5da6884da",
"source/SkillFactory/SkillFactory/tasks/manager_based/loaded_leg/env.py": "f18f82fe0fa50ef5591a247431d9230440021635c7025f0c5f458ef266ded855",
"source/SkillFactory/SkillFactory/tasks/manager_based/leg_skill/mdp.py": "3ac1792858de1f7576556661e4cfb8eb55c348e7fa3cf61c8e62d7fe04969bc0",
"source/SkillFactory/SkillFactory/legs/urdf.py": "694a6e0cf1c0825f02788982f528481932fe5cffb80738f130497c1040125c5a",
"source/SkillFactory/SkillFactory/tasks/manager_based/loaded_leg/__init__.py": "cabfbd1a302f26e808bf61b2d672eae8f29fd31fbeb2c4e559d1e5dc2631d28c",
"source/SkillFactory/SkillFactory/legs/support.py": "2abb489659001ba12e23becdbb02bec45f69a2e0b4d7efaeaec1c72e25d888cc",
"source/SkillFactory/SkillFactory/legs/tokens.py": "86dccd0a4af945700ff3382d2638d9ca8ebe06151c403d90b4cdb00a8fde7ad5",
"source/SkillFactory/SkillFactory/tasks/manager_based/leg_skill/__init__.py": "85417fb746517b22ec450803a02b09e23bf06421d0e78d83f272cff7de2eb678"
},
"status": "completed"
}
@@ -0,0 +1,42 @@
{
"passed": true,
"status": "completed",
"command": [
"/home/cen/IsaacLab/isaaclab.sh",
"-p",
"scripts/evaluate_standing.py",
"--package",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/candidate",
"--manifest",
"7cb478a3b17aaf307d2d6e3c2b49cc25aa9af1ee9d118dd731b52f618c27e628",
"--mode",
"full",
"--controller",
"policy",
"--output",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/full",
"--headless"
],
"pid": 346320,
"cleanup": {
"before": [
{
"pid": 346404,
"state": "Z",
"pgid": 346320,
"name": "Omniverse Hub"
},
{
"pid": 346413,
"state": "S",
"pgid": 346413,
"name": "omni.telemetry."
}
],
"remaining": [],
"clean": true
},
"returncode": 0,
"wall_seconds": 39.903371239997796,
"critical_log_lines": []
}
File diff suppressed because it is too large Load Diff
@@ -0,0 +1,44 @@
{
"passed": false,
"status": "failed",
"command": [
"/home/cen/IsaacLab/isaaclab.sh",
"-p",
"scripts/evaluate_standing.py",
"--package",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/candidate",
"--manifest",
"7cb478a3b17aaf307d2d6e3c2b49cc25aa9af1ee9d118dd731b52f618c27e628",
"--mode",
"holdout",
"--controller",
"hold",
"--output",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/hold",
"--headless"
],
"pid": 346685,
"cleanup": {
"before": [
{
"pid": 346769,
"state": "Z",
"pgid": 346685,
"name": "Omniverse Hub"
},
{
"pid": 346781,
"state": "S",
"pgid": 346781,
"name": "omni.telemetry."
}
],
"remaining": [],
"clean": true
},
"returncode": 1,
"wall_seconds": 39.833847160000005,
"critical_log_lines": [
"[ERROR] Command failed with code 1: \"/home/cen/IsaacLab/_isaac_sim/python.sh scripts/evaluate_standing.py --package /home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/candidate --manifest 7cb478a3b17aaf307d2d6e3c2b49cc25aa9af1ee9d118dd731b52f618c27e628 --mode holdout --controller hold --output /home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/hold --headless\""
]
}
File diff suppressed because it is too large Load Diff
@@ -0,0 +1,42 @@
{
"passed": true,
"status": "completed",
"command": [
"/home/cen/IsaacLab/isaaclab.sh",
"-p",
"scripts/evaluate_standing.py",
"--package",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/candidate",
"--manifest",
"7cb478a3b17aaf307d2d6e3c2b49cc25aa9af1ee9d118dd731b52f618c27e628",
"--mode",
"holdout",
"--controller",
"policy",
"--output",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/holdout",
"--headless"
],
"pid": 346496,
"cleanup": {
"before": [
{
"pid": 346587,
"state": "Z",
"pgid": 346496,
"name": "Omniverse Hub"
},
{
"pid": 346598,
"state": "S",
"pgid": 346598,
"name": "omni.telemetry."
}
],
"remaining": [],
"clean": true
},
"returncode": 0,
"wall_seconds": 40.58820937499695,
"critical_log_lines": []
}
@@ -0,0 +1,24 @@
{
"passed": true,
"package_only": true,
"isolated": true,
"device": "cpu",
"python": "3.12.3 (main, Aug 31 2026, 10:18:26) [GCC 13.3.0]",
"torch": "2.10.0+cpu",
"manifest_sha256": "7cb478a3b17aaf307d2d6e3c2b49cc25aa9af1ee9d118dd731b52f618c27e628",
"actor_sha256": "1384b5aa780cded1f4ed21b617ff8c883a5f82826c4cb4c9415b75f315b1e664",
"forbidden_imports": [],
"deterministic_batches": [
1,
16,
64,
256
],
"calls_per_batch": 100,
"input_state_unchanged": true,
"requires_grad": false,
"invalid_inputs_binding_corruption_rejected": true,
"status": "completed",
"package": "/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/candidate",
"stderr": "/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/stage2/inference_venv/lib/python3.12/site-packages/torch/_subclasses/functional_tensor.py:283: UserWarning: Failed to initialize NumPy: No module named 'numpy' (Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:84.)\n cpu = _conversion_method_template(device=torch.device(\"cpu\"))\n"
}
@@ -0,0 +1,41 @@
{
"format": "skillfactory.standing-deviation.v1",
"authorization": "explicit_user_instruction",
"operator_statement": "这个120秒耐久的飘逸超过的很小,可以视为通过",
"recorded_at": "2026-10-12T16:58:57.445471+08:00",
"approved": true,
"scope": "exact_candidate_and_trial_only",
"waived_gate": "stability.foot_drift",
"original_limit_m": 0.03,
"candidate_manifest_sha256": "7cb478a3b17aaf307d2d6e3c2b49cc25aa9af1ee9d118dd731b52f618c27e628",
"actor_sha256": "1384b5aa780cded1f4ed21b617ff8c883a5f82826c4cb4c9415b75f315b1e664",
"checkpoint_sha256": "f2197339dea3622ed92b82fdddff3400f2a82f3ca4d4fd3be6c72b45f1a77704",
"duration_s": 120.0,
"num_envs": 26,
"physical_samples": 60000,
"observed_failures": [
{
"environment": 6,
"max_foot_drift_m": 0.030620924223339147
},
{
"environment": 19,
"max_foot_drift_m": 0.03078258298377699
}
],
"evidence_sha256": {
"stability/result.json": "95a0f18edac40ece69430041b5083a011fe612a4770e899575ca985ba496e0b4",
"stability/metadata.json": "4c26c6fc3f2eb59b0fe6f2a5d1962c283f2da8e2c4047a33929f6643b17ab447",
"stability/physics_trace.pt": "3b35f32f71e87c812a24ae93101f05c917707691fea15ee523cca161846fe498",
"stability/memory.json": "3360e264f6062eff14fd1a015e670c26bdc1f84d1b7acc9dd568c35fd63868c5",
"stability_process.json": "e2af5a4fd4341dcb4ccecdef54dfa9c6404133a54cf2d7d05179056095ddd26d",
"stability.log": "e7678de92de394288d0fa5e570de090cb323bbd23411a49a7390f6ef30c6f380"
},
"not_authorized": [
"changing_original_results",
"changing_global_threshold",
"waiving_other_gates",
"new_weights_or_trials",
"hardware_safety_claims"
]
}
@@ -0,0 +1,34 @@
[WARNING] _isaac_sim is present but _isaac_sim/setup_conda_env.sh is missing; Isaac Sim env vars not exported.
[WARNING] Re-extract the Isaac Sim binary zip if you intend to use the bundled binary.
[INFO] Using Python: "/home/cen/IsaacLab/_isaac_sim/python.sh"
...sss.................................................................. [ 27%]
........................................................................ [ 55%]
........................................................................ [ 83%]
........................................... [100%]
=============================== warnings summary ===============================
../../../IsaacLab/source/isaaclab/isaaclab/utils/configclass.py:513
/home/cen/IsaacLab/source/isaaclab/isaaclab/utils/configclass.py:513: DeprecationWarning: 'RigidBodyMaterialCfg' is deprecated and will be removed in 5.0. Use 'isaaclab_physx.sim.spawners.materials.PhysxRigidBodyMaterialCfg' for PhysX properties, or 'isaaclab.sim.spawners.materials.RigidBodyMaterialBaseCfg' for solver-common properties only.
f1(*args, **kwargs)
../../../IsaacLab/source/isaaclab/isaaclab/utils/configclass.py:513
../../../IsaacLab/source/isaaclab/isaaclab/utils/configclass.py:513
../../../IsaacLab/source/isaaclab/isaaclab/utils/configclass.py:513
../../../IsaacLab/source/isaaclab/isaaclab/utils/configclass.py:513
../../../IsaacLab/source/isaaclab/isaaclab/utils/configclass.py:513
../../../IsaacLab/source/isaaclab/isaaclab/utils/configclass.py:513
../../../IsaacLab/source/isaaclab/isaaclab/utils/configclass.py:513
/home/cen/IsaacLab/source/isaaclab/isaaclab/utils/configclass.py:513: DeprecationWarning: 'RigidBodyPropertiesCfg' is deprecated and will be removed in 5.0. Use 'isaaclab_physx.sim.schemas.PhysxRigidBodyPropertiesCfg' for PhysX properties, or 'isaaclab.sim.schemas.RigidBodyBaseCfg' for solver-common properties only.
f1(*args, **kwargs)
../../../IsaacLab/source/isaaclab/isaaclab/utils/configclass.py:513
../../../IsaacLab/source/isaaclab/isaaclab/utils/configclass.py:513
../../../IsaacLab/source/isaaclab/isaaclab/utils/configclass.py:513
../../../IsaacLab/source/isaaclab/isaaclab/utils/configclass.py:513
../../../IsaacLab/source/isaaclab/isaaclab/utils/configclass.py:513
../../../IsaacLab/source/isaaclab/isaaclab/utils/configclass.py:513
../../../IsaacLab/source/isaaclab/isaaclab/utils/configclass.py:513
/home/cen/IsaacLab/source/isaaclab/isaaclab/utils/configclass.py:513: DeprecationWarning: 'ArticulationRootPropertiesCfg' is deprecated and will be removed in 5.0. Use 'isaaclab_physx.sim.schemas.PhysxArticulationRootPropertiesCfg' for PhysX properties, or 'isaaclab.sim.schemas.ArticulationRootBaseCfg' for solver-common properties only.
f1(*args, **kwargs)
-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
256 passed, 3 skipped, 15 warnings in 9.27s
@@ -0,0 +1,10 @@
## Review
- **Correct:** The prior P1 is closed. `load_waiver` hashes and parses the same immutable byte buffer (`scripts/standing_deviation.py:21–24`). Publication checks the copied waiver against the approved pin (`scripts/publish_standing.py:66–68,116`).
- **Correct:** Deterministic regressions cover replacement after reading and substitution/restoration during copying (`tests/test_standing_deviation.py:96–137`). Supplied logs report **256 passed, 3 skipped** and Ruff clean.
No issues found.
**Merge verdict: OK**
Read-only recheck; no commands, edits, simulations, or publication performed. This verdict confirms code-review closure, not operator authorization or completed publication.
@@ -0,0 +1,4 @@
[WARNING] _isaac_sim is present but _isaac_sim/setup_conda_env.sh is missing; Isaac Sim env vars not exported.
[WARNING] Re-extract the Isaac Sim binary zip if you intend to use the bundled binary.
[INFO] Using Python: "/home/cen/IsaacLab/_isaac_sim/python.sh"
All checks passed!
@@ -0,0 +1,782 @@
{
"passed": false,
"gates": {
"finite_complete_trace": true,
"floating_root": true,
"pgs_compatible": true,
"no_autoreset": true,
"no_fall": true,
"applied_load_matches_protocol": true,
"quaternion_norm": true,
"no_vertical_guide": true,
"height": true,
"tilt": true,
"foot_drift": false,
"contact": true,
"vertical_balance": true,
"limits": true,
"sensor_static_force": true,
"sensor_static_torque": true,
"memory": true
},
"physical_samples": 60000,
"mean_height_error_m": 0.0017030573373548721,
"rows": [
{
"mean_height_error_m": 0.0005739028682376856,
"max_height_error_m": 0.0006516454925156734,
"max_height_error_all_m": 0.0042078076114274165,
"max_tilt_rad": 0.01328328893823263,
"max_foot_drift_m": 0.0003817086663561282,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 7.698601708111601e-07,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559167436733368,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0024335683746997162,
"max_height_error_m": 0.005491036008796801,
"max_height_error_all_m": 0.005491036008796801,
"max_tilt_rad": 0.028170682477637952,
"max_foot_drift_m": 0.01365623688370796,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 8.788528662502566e-07,
"max_normalized_torque": 0.2657021867550174,
"max_normalized_velocity": 0.5559167436733368,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.001752732058525405,
"max_height_error_m": 0.003494131397209277,
"max_height_error_all_m": 0.004209208320579638,
"max_tilt_rad": 0.023651968877395332,
"max_foot_drift_m": 0.014317884731330726,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 6.34212829768012e-07,
"max_normalized_torque": 0.2657021867550174,
"max_normalized_velocity": 0.5559167436733368,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0014571713808471003,
"max_height_error_m": 0.0035820184459305904,
"max_height_error_all_m": 0.004212009738884082,
"max_tilt_rad": 0.025525089799726782,
"max_foot_drift_m": 0.01451681178214177,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 8.235070484761806e-07,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559167436733368,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0028766458701343987,
"max_height_error_m": 0.0070001660098649165,
"max_height_error_all_m": 0.0070001660098649165,
"max_tilt_rad": 0.03802712306318162,
"max_foot_drift_m": 0.01195365520153512,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 7.228121281715633e-07,
"max_normalized_torque": 0.2657021867550174,
"max_normalized_velocity": 0.5559167436733368,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0014501368551160612,
"max_height_error_m": 0.003767806123695483,
"max_height_error_all_m": 0.00421138389011394,
"max_tilt_rad": 0.026801996084162272,
"max_foot_drift_m": 0.016762926540175905,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 9.923693831389484e-07,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559163792118146,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0077213354424935324,
"max_height_error_m": 0.020213353465995898,
"max_height_error_all_m": 0.020213353465995898,
"max_tilt_rad": 0.05783302127961132,
"max_foot_drift_m": 0.030620924223339147,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 5.12516559090556e-07,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559163792118146,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0007149067874378461,
"max_height_error_m": 0.0009549139251328609,
"max_height_error_all_m": 0.004211741517982592,
"max_tilt_rad": 0.023567990896822353,
"max_foot_drift_m": 0.0017977309623659053,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 7.927707879645094e-07,
"max_normalized_torque": 0.2657021867550174,
"max_normalized_velocity": 0.5559163792118146,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.000787749776261723,
"max_height_error_m": 0.0010974882354355953,
"max_height_error_all_m": 0.004209685157737841,
"max_tilt_rad": 0.014638878611784931,
"max_foot_drift_m": 0.0017566646284406838,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 7.170597053372278e-07,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559164399554016,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0008133462863748381,
"max_height_error_m": 0.0012532649745560787,
"max_height_error_all_m": 0.004209416936836352,
"max_tilt_rad": 0.020645084435072736,
"max_foot_drift_m": 0.00040618808723046085,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 7.761654525594284e-07,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559163792118146,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.00039864706074153075,
"max_height_error_m": 0.0006657121886826656,
"max_height_error_all_m": 0.004209625553093066,
"max_tilt_rad": 0.01340308244678443,
"max_foot_drift_m": 0.0004603690674359409,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 8.232269496099708e-07,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559165006989887,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0006033583990005581,
"max_height_error_m": 0.0006630895843125484,
"max_height_error_all_m": 0.004213738273582568,
"max_tilt_rad": 0.013306528320339418,
"max_foot_drift_m": 0.0003727162844194601,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 7.554479237467523e-07,
"max_normalized_torque": 0.2657021867550174,
"max_normalized_velocity": 0.5559164399554016,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0005552422936645994,
"max_height_error_m": 0.0006548641433335445,
"max_height_error_all_m": 0.004210430215797534,
"max_tilt_rad": 0.013297276306088346,
"max_foot_drift_m": 0.00037897173137127775,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 7.449876387715755e-07,
"max_normalized_torque": 0.2657021867550174,
"max_normalized_velocity": 0.5559165006989887,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0005762547877526754,
"max_height_error_m": 0.0006548939456559322,
"max_height_error_all_m": 0.004208463262519946,
"max_tilt_rad": 0.013296749269705851,
"max_foot_drift_m": 0.0003799765063140337,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 7.580140694107488e-07,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559165006989887,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.002453487908354189,
"max_height_error_m": 0.005505609344444384,
"max_height_error_all_m": 0.005505609344444384,
"max_tilt_rad": 0.028278612412646138,
"max_foot_drift_m": 0.013841855059246492,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 9.031309527826232e-07,
"max_normalized_torque": 0.2657021867550174,
"max_normalized_velocity": 0.5559164399554016,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0017621707449714363,
"max_height_error_m": 0.003502446245155444,
"max_height_error_all_m": 0.004210907052955737,
"max_tilt_rad": 0.02366243376842146,
"max_foot_drift_m": 0.014326027828324121,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 6.246845767569706e-07,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559165006989887,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.001447578672357355,
"max_height_error_m": 0.00356172306438457,
"max_height_error_all_m": 0.00421409590145122,
"max_tilt_rad": 0.025444985939831456,
"max_foot_drift_m": 0.014409141183721125,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 9.110312076862622e-07,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559165006989887,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0028772112692014547,
"max_height_error_m": 0.007001000474891772,
"max_height_error_all_m": 0.007001000474891772,
"max_tilt_rad": 0.03803923485338886,
"max_foot_drift_m": 0.011986449202769811,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 7.022487149503722e-07,
"max_normalized_torque": 0.2657021867550174,
"max_normalized_velocity": 0.5559165006989887,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0014487033566726525,
"max_height_error_m": 0.003766733240089526,
"max_height_error_all_m": 0.004209714960060229,
"max_tilt_rad": 0.026726505987010665,
"max_foot_drift_m": 0.01672456715319223,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 1.055101138353225e-06,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559165006989887,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.00773894970051797,
"max_height_error_m": 0.020278918575248828,
"max_height_error_all_m": 0.020278918575248828,
"max_tilt_rad": 0.057941022266468155,
"max_foot_drift_m": 0.03078258298377699,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 5.166000419189637e-07,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559165006989887,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0006803565072078354,
"max_height_error_m": 0.0008932529201127193,
"max_height_error_all_m": 0.004209565948448291,
"max_tilt_rad": 0.023251757080382657,
"max_foot_drift_m": 0.0014582453890264776,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 7.113598774583283e-07,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559163792118146,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0007886898619596843,
"max_height_error_m": 0.001095938514671435,
"max_height_error_all_m": 0.004213142227134814,
"max_tilt_rad": 0.014629070828832023,
"max_foot_drift_m": 0.001759075882412823,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 7.865780350800203e-07,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559163792118146,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0008197085950206659,
"max_height_error_m": 0.001260864566764941,
"max_height_error_all_m": 0.0042129336108781,
"max_tilt_rad": 0.020673672798972736,
"max_foot_drift_m": 0.0003990844183059047,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 8.046230050917999e-07,
"max_normalized_torque": 0.2657021867550174,
"max_normalized_velocity": 0.5559163792118146,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.00039922315711737514,
"max_height_error_m": 0.0006688414325333736,
"max_height_error_all_m": 0.004209267925224414,
"max_tilt_rad": 0.01334071546404807,
"max_foot_drift_m": 0.0004673676174849101,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 8.186401782430033e-07,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559164399554016,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0006003791502217585,
"max_height_error_m": 0.0006603179683304927,
"max_height_error_all_m": 0.004208850692710986,
"max_tilt_rad": 0.013292606855339855,
"max_foot_drift_m": 0.00037885072321107633,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 7.226763597607158e-07,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559163792118146,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
},
{
"mean_height_error_m": 0.0005480336063366249,
"max_height_error_m": 0.0006480096091843746,
"max_height_error_all_m": 0.004208731483421435,
"max_tilt_rad": 0.01328441545112932,
"max_foot_drift_m": 0.0003783479748874635,
"contact_fraction": 1.0,
"vertical_balance_relative_error": 7.317862991770643e-07,
"max_normalized_torque": 0.2657021465155961,
"max_normalized_velocity": 0.5559167436733368,
"max_joint_limit_excess_rad": 0.0,
"max_guide_fz_N": 0.0,
"fallen": false
}
],
"load_reconstruction_max_error": 0.0,
"calibration_counts": [
500,
500,
500,
500,
500,
500,
500,
500,
500,
500,
500,
500,
500,
500,
500,
500,
500,
500,
500,
500,
500,
500,
500,
500,
500,
500
],
"sensor_static_max_force_error_N": 0.02974697828292827,
"sensor_static_max_torque_error_Nm": 0.0018565384255236957,
"memory_growth_fraction": 0.0,
"mode": "stability",
"controller": "policy",
"input_hash": "542eb98508f386705a3fd701be2b2cb791373822d522af9f52f654ba4afa44a7",
"metadata": {
"inputs": {
"mode": "stability",
"count": 26,
"duration_s": 120.0,
"seed": 85000,
"basis": [
[
0.0,
0.0,
0.0,
0.0,
0.0,
0.0
],
[
1.0,
0.0,
0.0,
0.0,
0.0,
0.0
],
[
-1.0,
0.0,
0.0,
0.0,
0.0,
0.0
],
[
0.0,
1.0,
0.0,
0.0,
0.0,
0.0
],
[
0.0,
-1.0,
0.0,
0.0,
0.0,
0.0
],
[
0.0,
0.0,
1.0,
0.0,
0.0,
0.0
],
[
0.0,
0.0,
-1.0,
0.0,
0.0,
0.0
],
[
0.0,
0.0,
0.0,
1.0,
0.0,
0.0
],
[
0.0,
0.0,
0.0,
-1.0,
0.0,
0.0
],
[
0.0,
0.0,
0.0,
0.0,
1.0,
0.0
],
[
0.0,
0.0,
0.0,
0.0,
-1.0,
0.0
],
[
0.0,
0.0,
0.0,
0.0,
0.0,
1.0
],
[
0.0,
0.0,
0.0,
0.0,
0.0,
-1.0
],
[
0.0,
0.0,
0.0,
0.0,
0.0,
0.0
],
[
1.0,
0.0,
0.0,
0.0,
0.0,
0.0
],
[
-1.0,
0.0,
0.0,
0.0,
0.0,
0.0
],
[
0.0,
1.0,
0.0,
0.0,
0.0,
0.0
],
[
0.0,
-1.0,
0.0,
0.0,
0.0,
0.0
],
[
0.0,
0.0,
1.0,
0.0,
0.0,
0.0
],
[
0.0,
0.0,
-1.0,
0.0,
0.0,
0.0
],
[
0.0,
0.0,
0.0,
1.0,
0.0,
0.0
],
[
0.0,
0.0,
0.0,
-1.0,
0.0,
0.0
],
[
0.0,
0.0,
0.0,
0.0,
1.0,
0.0
],
[
0.0,
0.0,
0.0,
0.0,
-1.0,
0.0
],
[
0.0,
0.0,
0.0,
0.0,
0.0,
1.0
],
[
0.0,
0.0,
0.0,
0.0,
0.0,
-1.0
]
],
"scale": [
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0,
1.0
],
"shift_s": [
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0
],
"period_s": 14.0
},
"input_hash": "542eb98508f386705a3fd701be2b2cb791373822d522af9f52f654ba4afa44a7",
"physics_dt": 0.002,
"control_dt": 0.01,
"height_m": 0.3345983503570176,
"floating_root": true,
"solver": "PGS",
"tgs_force_flag": false,
"autoresets": 0,
"total_mass_kg": 5.388350963592529,
"mass_kg": {
"virtual_payload": 3.0,
"anchor_base": 0.05000000074505806,
"FL_hip": 0.6779999732971191,
"FL_hip_rotor": 0.08900000154972076,
"FL_thigh": 1.1519999504089355,
"FL_thigh_rotor": 0.08900000154972076,
"FL_calf": 0.15399999916553497,
"FL_calf_rotor": 0.08900000154972076,
"FL_calflower": 0.02470862679183483,
"FL_foot": 0.03999999910593033,
"FL_calflower1": 0.02264302782714367
},
"limits": [
[
-1.0472,
1.0472
],
[
-1.5708,
3.4907
],
[
-2.7227,
-0.83776
]
],
"speeds": [
30.1,
30.1,
15.7
],
"efforts": [
23.700000762939453,
23.700000762939453,
45.43000030517578
]
},
"accepted_skill": false,
"package_manifest_sha256": "7cb478a3b17aaf307d2d6e3c2b49cc25aa9af1ee9d118dd731b52f618c27e628",
"actor_sha256": "1384b5aa780cded1f4ed21b617ff8c883a5f82826c4cb4c9415b75f315b1e664",
"physics_trace_sha256": "3b35f32f71e87c812a24ae93101f05c917707691fea15ee523cca161846fe498",
"source_hashes": {
"assets/generated/legs/go2_fl/manifest.json": "4cea30b0264444c2e0417d0c26d6ea9148698d3b77f559f12a9b33775a41e3eb",
"source/SkillFactory/SkillFactory/tasks/manager_based/loaded_leg/cfg.py": "36987e4eaf8973925ec7025fd29f458de3cef7f667fe794a989bbba8817df70f",
"source/SkillFactory/SkillFactory/tasks/manager_based/single_leg_anchor/mdp/__init__.py": "2b998fd206546dc11d0ee693ab1030a3d9949094a6ee25d07bdabefab84a3806",
"scripts/supervise_sim.py": "1d917d29f87cde38cd75343285ec18fdebc19bc7329c89d782b172ffb4779134",
"source/SkillFactory/SkillFactory/tasks/manager_based/leg_skill/leg_env.py": "abf20f5feb2114c02b5fc8647102430881640a0d3f3ac2a6503ff0dd4395cd98",
"assets/generated/support/go2_fl_3kg_v0/support.json": "9910845796906fc79094fbe9d48600291bb90d4d4abb1b02d696747f8ecccae0",
"source/SkillFactory/SkillFactory/tasks/manager_based/loaded_leg/env.py": "f18f82fe0fa50ef5591a247431d9230440021635c7025f0c5f458ef266ded855",
"scripts/runtime_common.py": "57934860bc6fe1df64128376737aa2a80aafe8f85cf4bee38d7f4373edfe908b",
"source/SkillFactory/SkillFactory/legs/__init__.py": "ca6f7541ea78f762be54147f3c978ed9564d045115fdfc975639e639696b8876",
"scripts/train_loaded_stand.py": "40de1c7b540b6a7df096508d7d533b679b7d8352b84dfe25b4dc201cec4f5652",
"source/SkillFactory/SkillFactory/tasks/manager_based/leg_skill/leg_env_cfg.py": "1c748307cc08e1b8367bb92fdf13a45bc3204abdf1cf423571b06d071e1b4815",
"scripts/evaluate_standing.py": "692ac51e07969565381ff49d08c82524bf275bfc50d555543fb20420992b96ad",
"scripts/standing_protocol.py": "d8ba91c96006687ba5d3e54fd986c915a034f16f11748ea06d717632b6322db1",
"source/SkillFactory/SkillFactory/tasks/manager_based/loaded_leg/__init__.py": "cabfbd1a302f26e808bf61b2d672eae8f29fd31fbeb2c4e559d1e5dc2631d28c",
"source/SkillFactory/SkillFactory/tasks/manager_based/single_leg_anchor/mdp/actions.py": "91aa6bd581437c932c19415c7c1abc45b2329b9e72a10d17aebeda6142a249da",
"configs/standing/go2_fl_support_v0.json": "bb2671ba1f1750baf6f0ec3ebb2f555d60ee252358867aa378fc85e07790c99a",
"source/SkillFactory/SkillFactory/legs/frozen_policy.py": "562d4ff6e4b5d1d569e16badcc44c972d303284e06371be22060a682aae38ffb",
"source/SkillFactory/SkillFactory/legs/policy.py": "cdfcc1da937d1b33652601d52a76b7c253ef64179822133317b0918ea84c2f6a",
"source/SkillFactory/SkillFactory/legs/resources/ground.usda": "26e7ef57327a36b59aacd0cc1ba0704ec4d86387450b4cab1f4020ae28bac70d",
"source/SkillFactory/SkillFactory/legs/urdf.py": "694a6e0cf1c0825f02788982f528481932fe5cffb80738f130497c1040125c5a",
"source/SkillFactory/SkillFactory/tasks/manager_based/single_leg_anchor/mdp/observations.py": "019a971445b99708386a6ffb27ec837f3b0aef1aa8fea59e8ae638783769b1f3",
"source/SkillFactory/SkillFactory/legs/spec.py": "e11d0df51131162c2416b099c2263e554ca72f13000a04828c79afc85a63f01d",
"source/SkillFactory/SkillFactory/legs/tokens.py": "86dccd0a4af945700ff3382d2638d9ca8ebe06151c403d90b4cdb00a8fde7ad5",
"source/SkillFactory/SkillFactory/tasks/manager_based/loaded_leg/mdp.py": "83ec027cf39675749c25ed0fe01518f8c430a768a5465e4c68efdfd5da6884da",
"source/SkillFactory/SkillFactory/legs/standing_ppo.py": "69d99ed40369a07ec1c8fee4220d9adbeb38958499d94c1bd7da054b78598565",
"source/SkillFactory/SkillFactory/legs/offline.py": "12ac018a2c76d046c8ed50ee105fcd6a33931f7c7a3bb027b6db332726ab064d",
"source/SkillFactory/SkillFactory/tasks/manager_based/leg_skill/__init__.py": "85417fb746517b22ec450803a02b09e23bf06421d0e78d83f272cff7de2eb678",
"source/SkillFactory/SkillFactory/tasks/manager_based/leg_skill/mdp.py": "3ac1792858de1f7576556661e4cfb8eb55c348e7fa3cf61c8e62d7fe04969bc0",
"source/SkillFactory/SkillFactory/legs/support.py": "2abb489659001ba12e23becdbb02bec45f69a2e0b4d7efaeaec1c72e25d888cc"
},
"status": "failed"
}
@@ -0,0 +1,44 @@
{
"passed": false,
"status": "failed",
"command": [
"/home/cen/IsaacLab/isaaclab.sh",
"-p",
"scripts/evaluate_standing.py",
"--package",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/candidate",
"--manifest",
"7cb478a3b17aaf307d2d6e3c2b49cc25aa9af1ee9d118dd731b52f618c27e628",
"--mode",
"stability",
"--controller",
"policy",
"--output",
"/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/stability",
"--headless"
],
"pid": 346906,
"cleanup": {
"before": [
{
"pid": 346990,
"state": "Z",
"pgid": 346906,
"name": "Omniverse Hub"
},
{
"pid": 346999,
"state": "S",
"pgid": 346999,
"name": "omni.telemetry."
}
],
"remaining": [],
"clean": true
},
"returncode": 1,
"wall_seconds": 223.3756301530011,
"critical_log_lines": [
"[ERROR] Command failed with code 1: \"/home/cen/IsaacLab/_isaac_sim/python.sh scripts/evaluate_standing.py --package /home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/candidate --manifest 7cb478a3b17aaf307d2d6e3c2b49cc25aa9af1ee9d118dd731b52f618c27e628 --mode stability --controller policy --output /home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/loaded_stand/acceptance_v1/v2/stability --headless\""
]
}
@@ -0,0 +1,20 @@
{
"format": "skillfactory.standing-bundle.v1",
"contract": "go2-fl-loaded-stand-v0",
"version": "1.0.0",
"stage": "candidate",
"hidden": 128,
"checkpoint_sha256": "f2197339dea3622ed92b82fdddff3400f2a82f3ca4d4fd3be6c72b45f1a77704",
"files": {
"ASSET_LICENSE": "84aac59fd3246e3ddc49d1387644e8fcf43b0def4f0b9fe687f372e90446df2d",
"README.md": "2c8cb0ea4f057b5c07600f2054aa58eed35b5e45e253ce7eec9bc1a532fc57ea",
"actor.pt": "1384b5aa780cded1f4ed21b617ff8c883a5f82826c4cb4c9415b75f315b1e664",
"binding.json": "22cd42a44f29913b2f0bd04e685e77acdcc1ef7b0218f62733dc401d55b6d710",
"example.pt": "f21e71bbc4b0e862a672d13daf67c31971f365426dd98662fc658ede6916c798",
"frozen_standing.py": "c19026845ea5ec98b21ed641cd294cccd502cf73907c08f0fe71a9a63067d8c6",
"requirements.txt": "812ea7d306f227902500eb76d816fe9f65f75d201edc6e766118ed68d3fdfe0d",
"selection.json": "276302e9374337aed8d4d3ea35408c4a0dc9971ea37bcecda8ccc5208b6a296a",
"task.json": "bb2671ba1f1750baf6f0ec3ebb2f555d60ee252358867aa378fc85e07790c99a",
"training.json": "443187246b6a600a11566eb1f1d3dca260f7c6a64684a847af7577afc4e521c9"
}
}
Binary file not shown.
@@ -0,0 +1,132 @@
"""Package-only guided standing inference. Python stdlib + Torch, no simulator/project imports.
SHA256 is an integrity check, not a signature. Caller owns the validated 100Hz
clock, xyzw sensor frames, last action, driver limits and floating guided rig.
"""
import hashlib
import json
import sys
from pathlib import Path
import torch
from torch import nn
CONTRACT = "go2-fl-loaded-stand-v0"
SHAPES = {
"joint_state": (3, 3),
"base_state": (13,),
"mount_wrench": (6,),
"foot_state": (10,),
"guide_wrench": (6,),
"goal": (1,),
}
CORE = {
"actor.pt",
"frozen_standing.py",
"binding.json",
"task.json",
"training.json",
"selection.json",
"README.md",
"requirements.txt",
"example.pt",
"ASSET_LICENSE",
}
def file_hash(path):
return hashlib.sha256(Path(path).read_bytes()).hexdigest()
def checked_manifest(root, expected):
root = Path(root).resolve()
if file_hash(root / "manifest.json") != expected:
raise ValueError("Package manifest identity mismatch")
m = json.loads((root / "manifest.json").read_text())
if m["format"] != "skillfactory.standing-bundle.v1" or m["contract"] != CONTRACT:
raise ValueError("Unsupported standing package")
if not CORE.issubset(m["files"]):
raise ValueError("Missing integrity entries")
for name, digest in m["files"].items():
path = (root / name).resolve()
if not path.is_relative_to(root) or not path.is_file() or file_hash(path) != digest:
raise ValueError(f"Bundle integrity failure: {name}")
return m
class Actor(nn.Module):
def __init__(self, hidden=128):
super().__init__()
self.register_buffer("offset", torch.zeros(45))
self.register_buffer("scale", torch.ones(45))
self.actor = nn.Sequential(
nn.Linear(45, hidden), nn.Tanh(), nn.Linear(hidden, hidden), nn.Tanh(), nn.Linear(hidden, 3)
)
def forward(self, x):
return self.actor(((x - self.offset) / self.scale).clamp(-10, 10)).tanh()
class FrozenStanding:
def __init__(self, directory, *, manifest_sha256, task_sha256, support_sha256, device="cpu"):
root = Path(directory).resolve()
m = checked_manifest(root, manifest_sha256)
if sys.version_info[:2] != (3, 12) or torch.__version__.split("+")[0] != "2.10.0":
raise RuntimeError("Validated dependencies: Python3.12 / Torch2.10.0")
self.binding = b = json.loads((root / "binding.json").read_text())
if b["task_sha256"] != task_sha256 or b["support_sha256"] != support_sha256:
raise ValueError("Unsupported physical binding")
if file_hash(root / "task.json") != task_sha256:
raise ValueError("Task file/binding mismatch")
self._actor = Actor(m["hidden"]).to(device)
self.device = next(self._actor.parameters()).device
self._actor.load_state_dict(
torch.load(root / "actor.pt", map_location=self.device, weights_only=True), strict=True
)
self._actor.eval().requires_grad_(False)
if (
any(not torch.isfinite(x).all() for x in self._actor.state_dict().values())
or (self._actor.scale <= 0).any()
):
raise ValueError("Invalid actor/normalization")
self.version = m["version"]
@torch.inference_mode()
def act(self, observation):
if set(observation) != set(SHAPES):
raise ValueError("Require all and only standing observation fields")
goal = observation["goal"]
if not isinstance(goal, torch.Tensor) or goal.ndim != 2 or len(goal) < 1:
raise ValueError("Require nonempty batch")
for key, shape in SHAPES.items():
x = observation[key]
if (
not isinstance(x, torch.Tensor)
or x.shape != (len(goal), *shape)
or x.dtype != torch.float32
or x.device != self.device
or not torch.isfinite(x).all()
):
raise ValueError(f"Invalid standing input: {key}")
if (goal - self.binding["height_m"]).abs().max() > 1e-5:
raise ValueError("Unsupported height goal")
if observation["guide_wrench"][:, 2].abs().max() != 0:
raise ValueError("Vertical guide not permitted")
quat = observation["base_state"][:, 3:7]
if (quat.norm(dim=-1) - 1).abs().max() > 0.002:
raise ValueError("Require unit xyzw quaternion")
x = torch.cat([observation[k].reshape(len(goal), -1) for k in SHAPES], -1)
x[:, 11] -= goal[:, 0]
action = self._actor(x)
if not torch.isfinite(action).all():
raise RuntimeError("Non-finite actor output")
return action
@torch.inference_mode()
def position_targets(self, observation):
action = self.act(observation)
soft = action.new_tensor(self.binding["soft_limits"])
return (action.new_tensor(self.binding["nominal"]) + self.binding["action_scale_rad"] * action).clamp(
min=soft[:, 0], max=soft[:, 1]
)
@@ -0,0 +1,75 @@
{
"format": "skillfactory.standing-bundle.v1",
"contract": "go2-fl-loaded-stand-v0",
"version": "1.0.0",
"stage": "accepted_with_deviation",
"hidden": 128,
"checkpoint_sha256": "f2197339dea3622ed92b82fdddff3400f2a82f3ca4d4fd3be6c72b45f1a77704",
"files": {
"ASSET_LICENSE": "84aac59fd3246e3ddc49d1387644e8fcf43b0def4f0b9fe687f372e90446df2d",
"DEVIATIONS.md": "f40548bfd46f8c098ec2e358780492aa329f0647b43fedc729b64139f3f051e0",
"README.md": "2c8cb0ea4f057b5c07600f2054aa58eed35b5e45e253ce7eec9bc1a532fc57ea",
"actor.pt": "1384b5aa780cded1f4ed21b617ff8c883a5f82826c4cb4c9415b75f315b1e664",
"binding.json": "22cd42a44f29913b2f0bd04e685e77acdcc1ef7b0218f62733dc401d55b6d710",
"evidence/acceptance_gate.json": "9349197540f38a4c722a651d8ddb82ee3ffa77142117d5a4e27e79f463622e3c",
"evidence/candidate_lock.json": "1dce2b41208f438bd98958d443dd8106574eb19063e66b9cfddf9edfbec258b4",
"evidence/engineering.json": "430ab38191962e1da218d4375c4a489019ac0d5e3e04333d236c41076a2b7cef",
"evidence/exploratory.json": "b7d6aac44a871ce33c0bfa6b5c8578749d62b882a516c54bda36afd3d3cce5e8",
"evidence/exploratory_process.json": "350ec331451ddb9b41d3386fe4414214242cd799d8ae72c881a296ccf23f6e70",
"evidence/full.json": "449f62b55ada8c84c6d67783c2891684984b1a6ed5ec5aff69aa08ee14b083df",
"evidence/full_process.json": "26712c9dc6c0e17bc5614ad2b7d39906a3e111c2d00d413147732089d4308767",
"evidence/hold.json": "aeb05e1fb283d5bbcad54ceec968031907b60d5c424fc585c8da71d50d844d1a",
"evidence/hold_process.json": "3b08d09a04bcdef07ad198cf9654b74edbb0d54eea6043031d52e7d852f04e78",
"evidence/holdout.json": "725ed47c750053f594236b1f611550a3e234fa82b546a84da3831266b589b8d5",
"evidence/holdout_process.json": "0750f7d89d7058b8873f54909f32429a0977dfd65c4e3b9623c93b620c79c58b",
"evidence/independent_candidate.json": "54396bb01be4c499b41a328547614980cf702232fd57c014ec6c37d96561fd3a",
"evidence/operator_waiver.json": "3c8c03efcedc3c6eb2a0a8ca8fdae2629834473fb122b2ce5bbcc39161344d1e",
"evidence/pytest_final.log": "911005ae3371eefae02b59c53e82d3fb9ce75ab4e873ee8d51ecaef3e0d93eff",
"evidence/review.md": "e04acdcc57bf3365fd05d5b224f384f94d2f55d60391af16efded1edaa932d41",
"evidence/ruff.log": "a5b097bbb606a1132331f167e0dab21849bbe45b37cc1957f17dcc665ece6218",
"evidence/stability.json": "95a0f18edac40ece69430041b5083a011fe612a4770e899575ca985ba496e0b4",
"evidence/stability_process.json": "e2af5a4fd4341dcb4ccecdef54dfa9c6404133a54cf2d7d05179056095ddd26d",
"evidence/tested_candidate_manifest.json": "7cb478a3b17aaf307d2d6e3c2b49cc25aa9af1ee9d118dd731b52f618c27e628",
"example.pt": "f21e71bbc4b0e862a672d13daf67c31971f365426dd98662fc658ede6916c798",
"frozen_standing.py": "c19026845ea5ec98b21ed641cd294cccd502cf73907c08f0fe71a9a63067d8c6",
"requirements.txt": "812ea7d306f227902500eb76d816fe9f65f75d201edc6e766118ed68d3fdfe0d",
"selection.json": "276302e9374337aed8d4d3ea35408c4a0dc9971ea37bcecda8ccc5208b6a296a",
"task.json": "bb2671ba1f1750baf6f0ec3ebb2f555d60ee252358867aa378fc85e07790c99a",
"training.json": "443187246b6a600a11566eb1f1d3dca260f7c6a64684a847af7577afc4e521c9"
},
"accepted_skill": true,
"acceptance_basis": "operator_waiver",
"unwaived_required_physics_passed": false,
"waiver_sha256": "3c8c03efcedc3c6eb2a0a8ca8fdae2629834473fb122b2ce5bbcc39161344d1e",
"waiver_receipts": [
{
"accepted_by_operator_waiver": true,
"raw_gate_passed": false,
"gate": "stability.foot_drift",
"original_limit_m": 0.03,
"observed_failures": [
{
"environment": 6,
"max_foot_drift_m": 0.030620924223339147
},
{
"environment": 19,
"max_foot_drift_m": 0.03078258298377699
}
]
}
],
"tested_candidate_manifest_sha256": "7cb478a3b17aaf307d2d6e3c2b49cc25aa9af1ee9d118dd731b52f618c27e628",
"tested_core_files": {
"ASSET_LICENSE": "84aac59fd3246e3ddc49d1387644e8fcf43b0def4f0b9fe687f372e90446df2d",
"README.md": "2c8cb0ea4f057b5c07600f2054aa58eed35b5e45e253ce7eec9bc1a532fc57ea",
"actor.pt": "1384b5aa780cded1f4ed21b617ff8c883a5f82826c4cb4c9415b75f315b1e664",
"binding.json": "22cd42a44f29913b2f0bd04e685e77acdcc1ef7b0218f62733dc401d55b6d710",
"example.pt": "f21e71bbc4b0e862a672d13daf67c31971f365426dd98662fc658ede6916c798",
"frozen_standing.py": "c19026845ea5ec98b21ed641cd294cccd502cf73907c08f0fe71a9a63067d8c6",
"requirements.txt": "812ea7d306f227902500eb76d816fe9f65f75d201edc6e766118ed68d3fdfe0d",
"selection.json": "276302e9374337aed8d4d3ea35408c4a0dc9971ea37bcecda8ccc5208b6a296a",
"task.json": "bb2671ba1f1750baf6f0ec3ebb2f555d60ee252358867aa378fc85e07790c99a",
"training.json": "443187246b6a600a11566eb1f1d3dca260f7c6a64684a847af7577afc4e521c9"
}
}
@@ -0,0 +1,2 @@
# Python3.12; CPU or CUDA PyTorch2.10.0 as separately validated
torch==2.10.0
@@ -0,0 +1,5 @@
{
"update": 1024,
"score": 0.000966311083175242,
"selection": "validation only, minimum penalized mean height error"
}
@@ -0,0 +1,28 @@
{
"schema_version": 1,
"task": "go2_fl_loaded_stand_v0",
"source_bundle": "assets/generated/legs/go2_fl",
"support_asset": "assets/generated/support/go2_fl_3kg_v0",
"payload_mass_kg": 3.0,
"payload_size_m": [0.18, 0.18, 0.08],
"payload_com_m": [0.0, 0.0, 0.06],
"sensor_carrier_mass_kg": 0.05,
"guide_stiffness": [250.0, 250.0, 0.0, 100.0, 100.0, 100.0],
"guide_damping": [20.0, 20.0, 0.0, 5.0, 5.0, 5.0],
"guide_wrench_limit": [60.0, 60.0, 0.0, 25.0, 25.0, 25.0],
"disturbance_amplitude": [5.0, 5.0, 10.0, 1.0, 1.0, 1.0],
"action_scale_rad": 0.5,
"initial_foot_clearance_m": 0.003,
"episode_seconds": 14.0,
"warmup_seconds": 2.0,
"height_error_limit_m": 0.06,
"tilt_limit_rad": 0.2617993877991494,
"foot_drift_limit_m": 0.03,
"contact_fraction_min": 0.98,
"vertical_balance_relative_tolerance": 0.10,
"sensor_static_force_tolerance_N": 1.5,
"sensor_static_torque_tolerance_Nm": 0.3,
"normalized_effort_velocity_limit": 1.00001,
"joint_limit_excess_rad": 0.01,
"seed": 71000
}
@@ -0,0 +1,42 @@
{
"format": "skillfactory.loaded-stand-ppo.v1",
"task_config": "configs/standing/go2_fl_support_v0.json",
"seed": 81000,
"validation_seed": 82000,
"holdout_seed": 83000,
"num_envs": 256,
"horizon": 32,
"updates": 2048,
"hidden": 128,
"initial_std": 0.15,
"learning_rate": 0.0003,
"gamma": 0.995,
"gae_lambda": 0.95,
"clip_ratio": 0.2,
"epochs": 4,
"minibatch_size": 2048,
"value_coefficient": 0.5,
"max_gradient_norm": 1.0,
"target_kl": 0.02,
"failure_penalty": 5.0,
"checkpoint_interval": 128,
"validation_interval": 256,
"train_amplitude_scale": [
0.5,
1.0
],
"train_time_shift_s": [
-0.5,
0.5
],
"validation_amplitude_scale": [
0.6,
1.0
],
"validation_time_shift_s": [
-0.25,
0.25
],
"max_wall_seconds": 5400,
"max_disk_bytes": 2147483648
}
+184
View File
@@ -0,0 +1,184 @@
{
"passed": true,
"milestone_complete": true,
"checks": {
"full_fixed_dataset": true,
"full_training": true,
"resumable": true,
"accepted_immutable_package": true,
"validation_only_selection": true,
"test": true,
"boundary": true,
"stability": true,
"independent_release_load": true,
"engineering": true
},
"scope": "Go2 FL only; fixed suspended anchor; 3-DoF foot-link Cartesian stationary reach; nominal reset; no contact, noise, morphology or whole-body transfer claims",
"release": "releases/go2_fl_reach-1.0.0",
"version": "1.0.0",
"config_sha256": "7e3e47a6d5b377dcdef98639dbb72cd585834c6e45b9750283f195bcb2cdcd52",
"actor_sha256": "cdfe459887e7629924df7e6a79fd2900eff184e5b25bc2a26bb585378f671c4a",
"manifest_sha256": "3106158d2ab214b09360528fb3c994a713435f05f801a8bde5115ebd16079655",
"selected_update": 20000,
"training_updates": 30000,
"transitions": 1024000,
"evaluation": {
"test": {
"max_normalized_torque": 0.4275040626525879,
"max_normalized_velocity": 0.8049103021621704,
"joint_limit_excess_rad": 0.0,
"anchor_pose_error": 0.0,
"invalid_states": 0,
"physical_samples": 8000,
"baseline_error_reduction": 0.9588127095617788,
"gates": {
"mean": true,
"p95": true,
"safety": true,
"success": true,
"baseline": true
},
"passed": true,
"weight_sha256": "cdfe459887e7629924df7e6a79fd2900eff184e5b25bc2a26bb585378f671c4a",
"package_manifest_sha256": "ca84aa6c5b6e87cdd2238bc6848c9f046d6b20eb71e4c66ea090631b81cc172b",
"seed": 31000,
"mode": "test",
"config_sha256": "7e3e47a6d5b377dcdef98639dbb72cd585834c6e45b9750283f195bcb2cdcd52",
"device": "cuda:0",
"asset_manifest_sha256": "4cea30b0264444c2e0417d0c26d6ea9148698d3b77f559f12a9b33775a41e3eb",
"wall_time_s": 21.81606545900013,
"status": "completed",
"policy": {
"episodes": 1024,
"mean_error_m": 0.0021836936939507723,
"p95_error_m": 0.004532407037913799,
"max_episode_mean_error_m": 0.006596561521291733,
"success_fraction": 1.0,
"success_wilson_95": [
0.9962624591182144,
1.0
]
},
"zero": {
"episodes": 1024,
"mean_error_m": 0.05301862955093384,
"p95_error_m": 0.08782369643449783,
"max_episode_mean_error_m": 0.10859119892120361,
"success_fraction": 0.0224609375,
"success_wilson_95": [
0.01501273728147908,
0.03347878125600763
]
}
},
"boundary": {
"max_normalized_torque": 0.40871959924697876,
"max_normalized_velocity": 0.7843982577323914,
"joint_limit_excess_rad": 0.0,
"anchor_pose_error": 0.0,
"invalid_states": 0,
"physical_samples": 2000,
"baseline_error_reduction": 0.968608317240211,
"gates": {
"mean": true,
"p95": true,
"safety": true
},
"passed": true,
"weight_sha256": "cdfe459887e7629924df7e6a79fd2900eff184e5b25bc2a26bb585378f671c4a",
"package_manifest_sha256": "ca84aa6c5b6e87cdd2238bc6848c9f046d6b20eb71e4c66ea090631b81cc172b",
"seed": 41000,
"mode": "boundary",
"config_sha256": "7e3e47a6d5b377dcdef98639dbb72cd585834c6e45b9750283f195bcb2cdcd52",
"device": "cuda:0",
"asset_manifest_sha256": "4cea30b0264444c2e0417d0c26d6ea9148698d3b77f559f12a9b33775a41e3eb",
"wall_time_s": 6.605768935000015,
"status": "completed",
"policy": {
"episodes": 256,
"mean_error_m": 0.002708765910938382,
"p95_error_m": 0.004719466902315617,
"max_episode_mean_error_m": 0.006831522565335035,
"success_fraction": 1.0,
"success_wilson_95": [
0.9852156082782743,
1.0
]
},
"zero": {
"episodes": 256,
"mean_error_m": 0.08628928661346436,
"p95_error_m": 0.12299373745918274,
"max_episode_mean_error_m": 0.12299486249685287,
"success_fraction": 0.0,
"success_wilson_95": [
0.0,
0.014784391721725852
]
}
},
"stability": {
"max_normalized_torque": 0.6637091636657715,
"max_normalized_velocity": 1.0000008344650269,
"joint_limit_excess_rad": 0.0,
"anchor_pose_error": 0.0,
"invalid_states": 0,
"physical_samples": 61000,
"control_steps": 12000,
"environment_control_steps": 768000,
"seconds_per_env": 120.0,
"num_envs": 64,
"device_used_before": 3544711168,
"device_used_max": 3544711168,
"memory_growth_fraction": 0.0,
"weight_sha256": "cdfe459887e7629924df7e6a79fd2900eff184e5b25bc2a26bb585378f671c4a",
"package_manifest_sha256": "ca84aa6c5b6e87cdd2238bc6848c9f046d6b20eb71e4c66ea090631b81cc172b",
"passed": true,
"config_sha256": "7e3e47a6d5b377dcdef98639dbb72cd585834c6e45b9750283f195bcb2cdcd52",
"device": "cuda:0",
"asset_manifest_sha256": "4cea30b0264444c2e0417d0c26d6ea9148698d3b77f559f12a9b33775a41e3eb",
"wall_time_s": 143.58381017600004,
"status": "completed"
}
},
"independent_load": {
"passed": true,
"actor_sha256": "cdfe459887e7629924df7e6a79fd2900eff184e5b25bc2a26bb585378f671c4a",
"manifest_sha256": "3106158d2ab214b09360528fb3c994a713435f05f801a8bde5115ebd16079655",
"python": "3.12.3 (main, Aug 31 2026, 10:18:26) [GCC 13.3.0]",
"torch": "2.10.0+cpu",
"device": "cpu",
"isolated": true,
"parameters_and_buffers_unchanged": true,
"requires_grad": false,
"training_or_simulator_imports": false,
"checks": [
"B=1:100 deterministic calls; no input mutation",
"B=16:100 deterministic calls; no input mutation",
"B=64:100 deterministic calls; no input mutation",
"B=256:100 deterministic calls; no input mutation",
"invalid active values/dtype/masks/context/hardware rejected",
"padding ignored",
"corruption rejected"
],
"package_only": true,
"stderr": "/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/stage2/inference_venv/lib/python3.12/site-packages/torch/_subclasses/functional_tensor.py:283: UserWarning: Failed to initialize NumPy: No module named 'numpy' (Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:84.)\n cpu = _conversion_method_template(device=torch.device(\"cpu\"))\n"
},
"engineering": {
"passed": true,
"pytest": "132 passed, 3 skipped (historical optional Kit integrations), 15 deprecation warnings",
"ruff": "all checks passed",
"git_diff_check": true,
"pytest_log_sha256": "b2c0981a6586f7d8ab1bc75f295c2ca48cc905d38756a82da2661abab539a605",
"ruff_log_sha256": "497df3a95b2cbe730aa38a8e0722c062245e1f3b990bd7890dd2a5cd01db6b6f",
"review_blockers_remaining": 0,
"regressions": [
"continuous vs resumed IQL exact update",
"invalid loss/hash/active inputs rejected",
"padding isolation",
"frozen params/buffers and caller inputs unchanged",
"final-only memory growth cannot pass",
"same weights but changed package cannot reuse evidence"
]
}
}
@@ -0,0 +1,23 @@
{
"passed": true,
"actor_sha256": "cdfe459887e7629924df7e6a79fd2900eff184e5b25bc2a26bb585378f671c4a",
"manifest_sha256": "3106158d2ab214b09360528fb3c994a713435f05f801a8bde5115ebd16079655",
"python": "3.12.3 (main, Aug 31 2026, 10:18:26) [GCC 13.3.0]",
"torch": "2.10.0+cpu",
"device": "cpu",
"isolated": true,
"parameters_and_buffers_unchanged": true,
"requires_grad": false,
"training_or_simulator_imports": false,
"checks": [
"B=1:100 deterministic calls; no input mutation",
"B=16:100 deterministic calls; no input mutation",
"B=64:100 deterministic calls; no input mutation",
"B=256:100 deterministic calls; no input mutation",
"invalid active values/dtype/masks/context/hardware rejected",
"padding ignored",
"corruption rejected"
],
"package_only": true,
"stderr": "/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/stage2/inference_venv/lib/python3.12/site-packages/torch/_subclasses/functional_tensor.py:283: UserWarning: Failed to initialize NumPy: No module named 'numpy' (Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:84.)\n cpu = _conversion_method_template(device=torch.device(\"cpu\"))\n"
}
Binary file not shown.
+29
View File
@@ -0,0 +1,29 @@
BSD 3-Clause License
Copyright (c) 2016-2022 HangZhou YuShu TECHNOLOGY CO.,LTD. ("Unitree Robotics")
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are met:
* Redistributions of source code must retain the above copyright notice, this
list of conditions and the following disclaimer.
* Redistributions in binary form must reproduce the above copyright notice,
this list of conditions and the following disclaimer in the documentation
and/or other materials provided with the distribution.
* Neither the name of the copyright holder nor the names of its
contributors may be used to endorse or promote products derived from
this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+34
View File
@@ -0,0 +1,34 @@
# Go2 FL Reach SkillBundle 1.0.0
仅固定悬空 Go2 FL,100Hz控制/500Hz物理、名义复位、2秒定点目标。
目标为 anchor frame 足端 link xyz/L,由 q_nominal+0.25*a,a∈[-.75,.75]^3 的FK生成;不是关节目标输入。
不支持接触承重、Go1、其他形态、移动根、噪声、步态、整机或实机。无 latent 技能库声明。
依赖:Python3.12,torch==2.10.0(其传递依赖由 pip 安装);CPU无Isaac依赖。
`python -m pip install torch==2.10.0 --index-url https://download.pytorch.org/whl/cpu`
只把本包目录加入 sys.path,然后:
```python
from frozen_policy import FrozenPolicy
import json, torch
from pathlib import Path
root = Path('/path/to/go2_fl_reach-1.0.0')
binding = json.loads((root/'binding.json').read_text())
p = FrozenPolicy(root, part_id='go2_fl', manifest_sha256=binding['manifest_sha256'])
tokens = torch.load(root/'example_tokens.pt', weights_only=True) # 仅接口示例;实际须提供实时状态/目标
command = p.act(tokens) # [B,8] float32,后5槽0;前3槽声明顺序、[-1,1]
q_target = p.position_targets(tokens) # [B,3] rad,已按soft_limits裁剪
```
实际调用方必须使用**自身资产的**manifest hash(示例从包读取仅供离线演示),按binding.joint_order映射。
输入完整旧leg-token-v1字典,字段/特征顺序见binding.feature_schema;mask bool,其余float32,同device。
归一化为固定物理尺度再经训练集mean/std;全部必要状态保存在actor.pt,推理无统计/参数更新。
joint_observation为(q-q_nominal)/scale、dq/URDF限速、last_applied_action;goal与foot位置均除length_scale_m。
静态tokens/context须匹配绑定;有效槽NaN/Inf、错shape/dtype/版本/硬件hash被拒绝,padding垃圾忽略。
执行器为显式IdealPD,Kp/Kd/effort/velocity/soft limits见binding;不能把输出直接当电机力矩。
调用方维持控制时序、有效传感、动作历史与驱动限矩限速。无实机watchdog/故障安全认证。
这是独立SkillBundle v1,不是声称通用SoftwareSpec v2 adapter已经实现。
完整性:manifest.json列出所有有效负载SHA256;外部SHA256SUMS固定manifest的信任锚。
Hash防意外损坏,不是签名;只执行可信来源的Python代码。拒绝覆盖已有发布目录。
任何权重、归一化、接口或元数据变化应发布新版本。查验manifest.status,candidate不是验收发布。
Binary file not shown.
+499
View File
@@ -0,0 +1,499 @@
{
"part_id": "go2_fl",
"manifest_sha256": "4cea30b0264444c2e0417d0c26d6ea9148698d3b77f559f12a9b33775a41e3eb",
"joint_order": [
"FL_hip_joint",
"FL_thigh_joint",
"FL_calf_joint"
],
"spec": {
"auxiliary_root_joints": [
"FL_hip_rotor_joint"
],
"decimation": 5,
"foot_body": "FL_foot",
"foot_collision_index": 0,
"joints": [
{
"action_scale": 0.25,
"kd": 0.5,
"kp": 25.0,
"name": "FL_hip_joint",
"q_default": 0.0
},
{
"action_scale": 0.25,
"kd": 0.5,
"kp": 25.0,
"name": "FL_thigh_joint",
"q_default": 0.8
},
{
"action_scale": 0.25,
"kd": 0.5,
"kp": 25.0,
"name": "FL_calf_joint",
"q_default": -1.5
}
],
"license_file": "../../LICENSE",
"license_id": "BSD-3-Clause",
"mount_joint": "FL_hip_joint",
"package_name": "go2_description",
"part_id": "go2_fl",
"physics_dt": 0.002,
"schema_version": 1,
"soft_limit_factor": 0.9,
"source_root": "/home/cen/Embodied_Workspace/unitree_ros/robots/go2_description",
"urdf": "urdf/go2_description.urdf"
},
"joints": [
{
"axis": [
1.0,
0.0,
0.0
],
"child": "FL_hip",
"limits": {
"effort": 23.7,
"lower": -1.0472,
"upper": 1.0472,
"velocity": 30.1
},
"name": "FL_hip_joint",
"nominal_axis": [
1.0,
0.0,
0.0
],
"nominal_position": [
0.0,
0.0,
0.0
],
"soft_limits": [
-0.94248,
0.94248
]
},
{
"axis": [
0.0,
1.0,
0.0
],
"child": "FL_thigh",
"limits": {
"effort": 23.7,
"lower": -1.5708,
"upper": 3.4907,
"velocity": 30.1
},
"name": "FL_thigh_joint",
"nominal_axis": [
0.0,
1.0,
0.0
],
"nominal_position": [
0.0,
0.0955,
0.0
],
"soft_limits": [
-1.3177249999999998,
3.237625
]
},
{
"axis": [
0.0,
1.0,
0.0
],
"child": "FL_calf",
"limits": {
"effort": 45.43,
"lower": -2.7227,
"upper": -0.83776,
"velocity": 15.7
},
"name": "FL_calf_joint",
"nominal_axis": [
0.0,
1.0,
0.0
],
"nominal_position": [
-0.15279684736159835,
0.0955,
-0.14839852909094622
],
"soft_limits": [
-2.6284530000000004,
-0.9320069999999999
]
}
],
"length_scale_m": 0.5455,
"source": {
"commit": "daadf41ee9afce8f90fdc09a98506012691fa122",
"license_id": "BSD-3-Clause",
"license_sha256": "84aac59fd3246e3ddc49d1387644e8fcf43b0def4f0b9fe687f372e90446df2d",
"sha256": "7d19fe48e2e689ee1a032ab99f2a4a8b671d87e73de48d3e65811682a5b48b9e",
"urdf": "/home/cen/Embodied_Workspace/unitree_ros/robots/go2_description/urdf/go2_description.urdf"
},
"feature_schema": {
"version": 1,
"joint_description": [
"pivot_x/L",
"pivot_y/L",
"pivot_z/L",
"axis_x",
"axis_y",
"axis_z",
"q_nominal/pi",
"control_min/pi",
"control_max/pi",
"effort/Emax",
"velocity/Vmax",
"kp*pi/Emax",
"kd*Vmax/Emax",
"action_scale/pi",
"child_mass/Mknown",
"child_Ixx/ML2",
"child_Ixy/ML2",
"child_Ixz/ML2",
"child_Iyy/ML2",
"child_Iyz/ML2",
"child_Izz/ML2",
"child_com_x/L",
"child_com_y/L",
"child_com_z/L",
"inertial_valid"
],
"joint_observation": [
"(q-q_nominal)/action_scale",
"dq/velocity_limit",
"last_applied_action"
],
"effector_description": [
"nominal_x/L",
"nominal_y/L",
"nominal_z/L",
"collision_x/L",
"collision_y/L",
"collision_z/L",
"collision_qx",
"collision_qy",
"collision_qz",
"collision_qw",
"sphere_radius/L",
"box_x/L",
"box_y/L",
"box_z/L",
"is_sphere",
"is_box"
],
"effector_observation": [
"link_x/L",
"link_y/L",
"link_z/L",
"link_vx/(L/s)",
"link_vy/(L/s)",
"link_vz/(L/s)",
"normal_Fx/(Mknown*g)",
"normal_Fy/(Mknown*g)",
"normal_Fz/(Mknown*g)",
"contact",
"selected_collision_gap/L"
],
"context": [
"gravity_x",
"gravity_y",
"gravity_z",
"control_dt_s",
"ground_z_relative_to_anchor/L",
"length_scale_m",
"known_mass_scale_kg",
"max_effort/(Mknown*g*L)",
"max_joint_velocity_rad_per_s",
"has_missing_source_inertia"
],
"goal": [
"target_link_x/L",
"target_link_y/L",
"target_link_z/L"
],
"force_semantics": "normal contact force only; no friction wrench",
"inertia_semantics": "Source child COM inertia and relative COM offset rotated into anchor nominal frame; not fixed-subtree aggregated",
"mass_semantics": "Known source mass excludes synthetic anchor; not a prediction of support force"
},
"max_joints": 8,
"context": [
0.0,
0.0,
-1.0,
0.009999999776482582,
-1.183318018913269,
0.5454999804496765,
2.2909998893737793,
3.7055602073669434,
30.100000381469727,
1.0
],
"static_tokens": {
"joint_description": [
[
0.0,
0.0,
0.0,
1.0,
0.0,
0.0,
0.0,
-0.07957746833562851,
0.07957746833562851,
0.5216817259788513,
1.0,
1.7288094758987427,
0.33127889037132263,
0.07957746833562851,
0.295940637588501,
0.0007040874916128814,
-4.415215244080173e-06,
1.6282023125313572e-06,
0.0012966945068910718,
-2.0829254481213866e-06,
0.0008742419886402786,
-0.00989917479455471,
0.0035563702695071697,
-0.000192483959835954,
1.0
],
[
0.0,
0.17506875097751617,
0.0,
0.0,
1.0,
0.0,
0.2546479105949402,
0.17507043480873108,
0.3342253863811493,
0.5216817259788513,
1.0,
1.7288094758987427,
0.33127889037132263,
0.07957746833562851,
0.5028371810913086,
0.004511878360062838,
0.0009393356740474701,
-0.0035138893872499466,
0.008507723920047283,
0.0007339900475926697,
0.005565374158322811,
-0.04777860268950462,
-0.040879927575588226,
-0.03684582561254501,
1.0
],
[
-0.28010421991348267,
0.17506875097751617,
-0.27204129099845886,
0.0,
1.0,
0.0,
-0.47746482491493225,
-0.5570423007011414,
-0.39788734912872314,
1.0,
0.5215947031974792,
1.7288094758987427,
0.33127889037132263,
0.07957746833562851,
0.0672195553779602,
0.0009218948544003069,
-7.442903097398812e-06,
0.000761082861572504,
0.0016135338228195906,
9.610687811800744e-06,
0.0007105613476596773,
0.14349472522735596,
-0.0017873510951176286,
-0.15476909279823303,
1.0
],
[
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0
],
[
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0
],
[
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0
],
[
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0
],
[
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0
]
],
"joint_mask": [
true,
true,
true,
false,
false,
false,
false,
false
],
"effector_description": [
[
-0.028558166697621346,
0.17506875097751617,
-0.5706872940063477,
-0.0036663610953837633,
0.0,
0.0,
0.0,
0.0,
0.0,
1.0,
0.04032997414469719,
0.0,
0.0,
0.0,
1.0,
0.0
]
],
"effector_mask": [
true
]
}
}
+79
View File
@@ -0,0 +1,79 @@
{
"experiment": "go2_fl_reach_iql_v1_attempt2",
"scope": "Go2 FL only; fixed suspended anchor; 3-DoF foot-link Cartesian stationary reach; nominal reset; no contact, noise, morphology or whole-body transfer claims",
"bundle": "assets/generated/legs/go2_fl",
"max_joints": 8,
"num_envs": 256,
"episode_steps": 200,
"settle_steps": 150,
"goal_action_range": 0.75,
"collection": {
"seed": 12000,
"shards": 20,
"noise_std_by_quartile": [
0.0,
0.03,
0.1,
0.25
],
"feedback_gain": 1.5,
"velocity_gain_s": 0.04,
"integral_gain_s_inv": 8.0
},
"training": {
"seed": 42,
"updates": 30000,
"batch_size": 512,
"hidden": 128,
"lr": 0.0003,
"discount": 0.99,
"expectile": 0.7,
"temperature": 3.0,
"max_weight": 100.0,
"target_tau": 0.005,
"checkpoint_every": 5000
},
"selection": {
"seed": 21000,
"batches": 2,
"candidate_updates": [
10000,
20000,
30000
],
"rule": "lowest final-window mean error among validation candidates satisfying all nominal gates; no selection on test"
},
"test": {
"seed": 31000,
"batches": 4,
"boundary_seed": 41000,
"boundary_batches": 1,
"stability_seed": 51000,
"stability_envs": 64,
"stability_seconds": 120
},
"acceptance": {
"mean_error_m": 0.008,
"p95_error_m": 0.015,
"success_radius_m": 0.015,
"success_fraction": 0.95,
"baseline_error_reduction": 0.6,
"boundary_mean_error_m": 0.012,
"boundary_p95_error_m": 0.02,
"max_normalized_torque": 1.00001,
"max_normalized_velocity": 1.00001,
"joint_limit_tolerance_rad": 0.01,
"anchor_pose_max_error": 0.0001,
"invalid_states": 0,
"unexpected_terminations": 0,
"memory_growth_fraction": 0.05
},
"budget": {
"gpu_count": 1,
"gpu": "RTX 5080 16 GiB",
"wall_seconds": 5400,
"disk_gb": 10,
"max_training_attempts": 2
},
"release_version": "1.0.0"
}
File diff suppressed because it is too large Load Diff
@@ -0,0 +1,215 @@
{
"config_sha256": "7e3e47a6d5b377dcdef98639dbb72cd585834c6e45b9750283f195bcb2cdcd52",
"asset_manifest_sha256": "4cea30b0264444c2e0417d0c26d6ea9148698d3b77f559f12a9b33775a41e3eb",
"feature_schema": {
"version": 1,
"joint_description": [
"pivot_x/L",
"pivot_y/L",
"pivot_z/L",
"axis_x",
"axis_y",
"axis_z",
"q_nominal/pi",
"control_min/pi",
"control_max/pi",
"effort/Emax",
"velocity/Vmax",
"kp*pi/Emax",
"kd*Vmax/Emax",
"action_scale/pi",
"child_mass/Mknown",
"child_Ixx/ML2",
"child_Ixy/ML2",
"child_Ixz/ML2",
"child_Iyy/ML2",
"child_Iyz/ML2",
"child_Izz/ML2",
"child_com_x/L",
"child_com_y/L",
"child_com_z/L",
"inertial_valid"
],
"joint_observation": [
"(q-q_nominal)/action_scale",
"dq/velocity_limit",
"last_applied_action"
],
"effector_description": [
"nominal_x/L",
"nominal_y/L",
"nominal_z/L",
"collision_x/L",
"collision_y/L",
"collision_z/L",
"collision_qx",
"collision_qy",
"collision_qz",
"collision_qw",
"sphere_radius/L",
"box_x/L",
"box_y/L",
"box_z/L",
"is_sphere",
"is_box"
],
"effector_observation": [
"link_x/L",
"link_y/L",
"link_z/L",
"link_vx/(L/s)",
"link_vy/(L/s)",
"link_vz/(L/s)",
"normal_Fx/(Mknown*g)",
"normal_Fy/(Mknown*g)",
"normal_Fz/(Mknown*g)",
"contact",
"selected_collision_gap/L"
],
"context": [
"gravity_x",
"gravity_y",
"gravity_z",
"control_dt_s",
"ground_z_relative_to_anchor/L",
"length_scale_m",
"known_mass_scale_kg",
"max_effort/(Mknown*g*L)",
"max_joint_velocity_rad_per_s",
"has_missing_source_inertia"
],
"goal": [
"target_link_x/L",
"target_link_y/L",
"target_link_z/L"
],
"force_semantics": "normal contact force only; no friction wrench",
"inertia_semantics": "Source child COM inertia and relative COM offset rotated into anchor nominal frame; not fixed-subtree aggregated",
"mass_semantics": "Known source mass excludes synthetic anchor; not a prediction of support force"
},
"shards": [
{
"path": "shard_000.pt",
"sha256": "6c22000bd871a41b88b90851002db418d025988f15c02b77a83305c3106ed9b7",
"seed": 12000,
"transitions": 51200
},
{
"path": "shard_001.pt",
"sha256": "53ae0894c392db2823eeac22551a951fe0f93418b0d8a8674803d78855c1d2d5",
"seed": 12001,
"transitions": 51200
},
{
"path": "shard_002.pt",
"sha256": "69103fc3729308af87096a75539e143d5f2ddba09ce250048a9dec6829b12775",
"seed": 12002,
"transitions": 51200
},
{
"path": "shard_003.pt",
"sha256": "a48b0354dfb69ebe2f18539c652f3137f5cec1f2aa9208aba7ca6fd60cefebbe",
"seed": 12003,
"transitions": 51200
},
{
"path": "shard_004.pt",
"sha256": "1025160d3b3de8025e36316e5876af8ea4666123fd10037848b4c14a7fe68acd",
"seed": 12004,
"transitions": 51200
},
{
"path": "shard_005.pt",
"sha256": "21a32b080efbb96e30f1da487ce596a19f2bb62dbe712bbf1ede0aba5252cd31",
"seed": 12005,
"transitions": 51200
},
{
"path": "shard_006.pt",
"sha256": "d55dee7a9b02b47c017c7f23cc340c417972f35722c3e4d4e6dc40d63aa7c42e",
"seed": 12006,
"transitions": 51200
},
{
"path": "shard_007.pt",
"sha256": "0eb5026af22bd86c78a9cd4ff079a450ceaeef525bf467f45d089c6891ff8885",
"seed": 12007,
"transitions": 51200
},
{
"path": "shard_008.pt",
"sha256": "8cde8a3f84d4b9ef68f055382badc18b91a480d223be62abbb49b92e1527ded4",
"seed": 12008,
"transitions": 51200
},
{
"path": "shard_009.pt",
"sha256": "3ff78a3d10dcfc4b2b98d156236251fcfd1fe79603923f08767406d623fa3911",
"seed": 12009,
"transitions": 51200
},
{
"path": "shard_010.pt",
"sha256": "8206486abff6fa604fd3f4c18ad107d00faa58bdcbd1269b38a5cbbdb3165517",
"seed": 12010,
"transitions": 51200
},
{
"path": "shard_011.pt",
"sha256": "c001d4561c129f08ed3dc62ed118e4bc9169efd0e5990a27b48225f4874fddda",
"seed": 12011,
"transitions": 51200
},
{
"path": "shard_012.pt",
"sha256": "f4c25b40b84b47d759e74f4b57f84772624431707fb1bb3da36f41606c7661a4",
"seed": 12012,
"transitions": 51200
},
{
"path": "shard_013.pt",
"sha256": "d7f4513cffe1fdcd3ffd2b4c7a9f2e32a204184e1256c3992f3aed68d6effa97",
"seed": 12013,
"transitions": 51200
},
{
"path": "shard_014.pt",
"sha256": "5f80cc08f31572ba74225f9acf8b80c17aee3d3347e90088cad390baa9a270a9",
"seed": 12014,
"transitions": 51200
},
{
"path": "shard_015.pt",
"sha256": "faf2523f065132f97e9e69c746da6404b0790c78cdf8d5aec9612e32d0e34a56",
"seed": 12015,
"transitions": 51200
},
{
"path": "shard_016.pt",
"sha256": "0a95607f01d6740fdcc2957757860b99f9a9ebdb47077918c9888f4001a7712e",
"seed": 12016,
"transitions": 51200
},
{
"path": "shard_017.pt",
"sha256": "68780da8bd6541343f71e35e6d4ff5525859a61e52a4456f1165848a3f099894",
"seed": 12017,
"transitions": 51200
},
{
"path": "shard_018.pt",
"sha256": "4d329c699320fca3432f6f4510a08dbff85c5311922fae901446611c39876b46",
"seed": 12018,
"transitions": 51200
},
{
"path": "shard_019.pt",
"sha256": "539b3a8c311878cffcffe534f7ace2bf12a6361013b7fc7c0f231b64313133d7",
"seed": 12019,
"transitions": 51200
}
],
"complete": true,
"transitions": 1024000,
"bootstrap": "No true terminations. Artificial 2s truncations bootstrap from pre-reset next-state."
}
@@ -0,0 +1,17 @@
{
"passed": true,
"pytest": "132 passed, 3 skipped (historical optional Kit integrations), 15 deprecation warnings",
"ruff": "all checks passed",
"git_diff_check": true,
"pytest_log_sha256": "b2c0981a6586f7d8ab1bc75f295c2ca48cc905d38756a82da2661abab539a605",
"ruff_log_sha256": "497df3a95b2cbe730aa38a8e0722c062245e1f3b990bd7890dd2a5cd01db6b6f",
"review_blockers_remaining": 0,
"regressions": [
"continuous vs resumed IQL exact update",
"invalid loss/hash/active inputs rejected",
"padding isolation",
"frozen params/buffers and caller inputs unchanged",
"final-only memory growth cannot pass",
"same weights but changed package cannot reuse evidence"
]
}
@@ -0,0 +1,10 @@
filelock==3.32.3
fsspec==2026.7.0
Jinja2==3.1.6
MarkupSafe==3.0.3
mpmath==1.3.0
networkx==3.6.1
setuptools==78.1.0
sympy==1.14.0
torch==2.10.0+cpu
typing_extensions==4.16.0
@@ -0,0 +1,31 @@
{
"git_base": "8019d6f8d6b4004ab09137b35faa87a3117f2b56",
"dirty_worktree": true,
"python": "3.12.13 (tags/v3.12.13-dirty:3bb231a, Apr 22 2026, 14:10:08) [GCC 11.4.0]",
"platform": "Linux-7.0.0-38-generic-x86_64-with-glibc2.39",
"versions": {
"torch": "2.10.0+cu128",
"isaaclab": "6.1.17",
"isaacsim": "not available as distribution"
},
"source_sha256": {
"scripts/verify_stage2_delivery.py": "506a43f47631c109538361a22734052ea1fb21982e9892e9f9355f2833faa723",
"scripts/stage2_sim.py": "d6dfd1a2d0bef0a647cda85f4d04227e24806f897d112a322f757031ab9d7f38",
"scripts/train_offline_leg.py": "cf05cd55c6d5d875417928bc9d31d54d1d7c073d2877563d749bdf70e556a2b5",
"scripts/export_leg_bundle.py": "75470eaa2e0ac3016b6ffa1febe1e34ad5e4b3b44fea798cd37dde9ffba6c0b0",
"scripts/verify_frozen_leg.py": "2320b8b8e3f06e402dab5ff605b91356027053438281ab18aaf428622a00580c",
"source/SkillFactory/SkillFactory/legs/offline.py": "12ac018a2c76d046c8ed50ee105fcd6a33931f7c7a3bb027b6db332726ab064d",
"source/SkillFactory/SkillFactory/legs/frozen_policy.py": "562d4ff6e4b5d1d569e16badcc44c972d303284e06371be22060a682aae38ffb",
"tests/test_offline_leg.py": "6d94bb7a91a0487bd3de2ecbf4a2caaaa2515af1f1289ed8f1142d33604b04b4",
"configs/training/go2_fl_offline_v1_attempt2.json": "1c56a59a74cb35810b3a43e409509c03b9f96bcb285a2e09b887e01cef9df114"
},
"data_collection_source_sha256": "c925a400d2a755a4791023748db819386e795a747034d501eac7f2e666b2e455",
"asset_manifest_sha256": "4cea30b0264444c2e0417d0c26d6ea9148698d3b77f559f12a9b33775a41e3eb",
"repeated_evaluation_reason": "Evidence-only rerun with whole-candidate manifest hash; same weights/goals/thresholds; no selection or tuning after test. First results retained in pre_manifest_evaluation/.",
"review_findings_resolved": [
"Final durability memory sample included + mocked tail-growth regression",
"Unindexed CUDA device canonicalized",
"Full candidate manifest hash bound to simulation and isolated-loader evidence + stale-package regression",
"Reproduction commands now capture mandatory logs with pipefail"
]
}
@@ -0,0 +1,7 @@
{
"passed": true,
"source_update": 5000,
"compared_update": 5001,
"full_learner_optimizer_rng_exact": true,
"device": "cuda:0"
}
@@ -0,0 +1,23 @@
{
"max_normalized_torque": 0.6637091636657715,
"max_normalized_velocity": 1.0000008344650269,
"joint_limit_excess_rad": 0.0,
"anchor_pose_error": 0.0,
"invalid_states": 0,
"physical_samples": 61000,
"control_steps": 12000,
"environment_control_steps": 768000,
"seconds_per_env": 120.0,
"num_envs": 64,
"device_used_before": 3544711168,
"device_used_max": 3544711168,
"memory_growth_fraction": 0.0,
"weight_sha256": "cdfe459887e7629924df7e6a79fd2900eff184e5b25bc2a26bb585378f671c4a",
"package_manifest_sha256": "ca84aa6c5b6e87cdd2238bc6848c9f046d6b20eb71e4c66ea090631b81cc172b",
"passed": true,
"config_sha256": "7e3e47a6d5b377dcdef98639dbb72cd585834c6e45b9750283f195bcb2cdcd52",
"device": "cuda:0",
"asset_manifest_sha256": "4cea30b0264444c2e0417d0c26d6ea9148698d3b77f559f12a9b33775a41e3eb",
"wall_time_s": 143.58381017600004,
"status": "completed"
}
@@ -0,0 +1,23 @@
{
"passed": true,
"actor_sha256": "cdfe459887e7629924df7e6a79fd2900eff184e5b25bc2a26bb585378f671c4a",
"manifest_sha256": "ca84aa6c5b6e87cdd2238bc6848c9f046d6b20eb71e4c66ea090631b81cc172b",
"python": "3.12.3 (main, Aug 31 2026, 10:18:26) [GCC 13.3.0]",
"torch": "2.10.0+cpu",
"device": "cpu",
"isolated": true,
"parameters_and_buffers_unchanged": true,
"requires_grad": false,
"training_or_simulator_imports": false,
"checks": [
"B=1:100 deterministic calls; no input mutation",
"B=16:100 deterministic calls; no input mutation",
"B=64:100 deterministic calls; no input mutation",
"B=256:100 deterministic calls; no input mutation",
"invalid active values/dtype/masks/context/hardware rejected",
"padding ignored",
"corruption rejected"
],
"package_only": true,
"stderr": "/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/stage2/inference_venv/lib/python3.12/site-packages/torch/_subclasses/functional_tensor.py:283: UserWarning: Failed to initialize NumPy: No module named 'numpy' (Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:84.)\n cpu = _conversion_method_template(device=torch.device(\"cpu\"))\n"
}
File diff suppressed because it is too large Load Diff
@@ -0,0 +1,24 @@
{
"format": "skillfactory.skillbundle.v1",
"interface": "go2-fl-reach-leg-token-v1",
"version": "1.0.0",
"status": "candidate",
"frozen": true,
"hidden": 128,
"dependencies": {
"python": "3.12.x",
"torch": "2.10.0 (CPU validated; CUDA evaluated separately)"
},
"scope": "Go2 FL only; fixed suspended anchor; 3-DoF foot-link Cartesian stationary reach; nominal reset; no contact, noise, morphology or whole-body transfer claims",
"files": {
"ASSET_LICENSE": "84aac59fd3246e3ddc49d1387644e8fcf43b0def4f0b9fe687f372e90446df2d",
"README.md": "7c6cfaea271595f106e4f042afb94a4e16916c75405d17425a02fd08acf37dde",
"actor.pt": "cdfe459887e7629924df7e6a79fd2900eff184e5b25bc2a26bb585378f671c4a",
"binding.json": "a8b6decc1e311b853943e088b48773a2de78b7ed1d67e465f53b2d849f242dbd",
"config.json": "1c56a59a74cb35810b3a43e409509c03b9f96bcb285a2e09b887e01cef9df114",
"example_tokens.pt": "f91ac1f14f9dc63d9f857f5cf901573ef21a219f6618e506bd0b92b57df9538f",
"frozen_policy.py": "562d4ff6e4b5d1d569e16badcc44c972d303284e06371be22060a682aae38ffb",
"requirements.txt": "69191f795f024589b482e6bf90c5db6a5631c0296f383f193eb1c1f76dfeaefd",
"selection.json": "cdf8a22ac773b88322556ec77a35e5247814a13258bd684778754e50413ff4c6"
}
}
@@ -0,0 +1,21 @@
{
"passed": true,
"training_complete": true,
"status": "completed",
"updates": 30000,
"elapsed_s": 134.28458627700002,
"identity": {
"config_sha256": "7e3e47a6d5b377dcdef98639dbb72cd585834c6e45b9750283f195bcb2cdcd52",
"dataset_sha256": "1228d9ee3b2cb65b6b98d9d193e7b66f8506a35436c6c5edea3c3816f5b60023",
"source_sha256": {
"scripts/train_offline_leg.py": "cf05cd55c6d5d875417928bc9d31d54d1d7c073d2877563d749bdf70e556a2b5",
"source/SkillFactory/SkillFactory/legs/offline.py": "12ac018a2c76d046c8ed50ee105fcd6a33931f7c7a3bb027b6db332726ab064d",
"source/SkillFactory/SkillFactory/legs/frozen_policy.py": "562d4ff6e4b5d1d569e16badcc44c972d303284e06371be22060a682aae38ffb"
}
},
"checkpoint_sha256": "5dce2d6a09e58a96aae13f2a63747de3f42bb79ace76ae8859b6d5e8dd674746",
"transitions": 1024000,
"device": "cuda:0",
"acceptance_complete": false,
"algorithm": "IQL; twin-Q TD, expectile V, advantage-weighted deterministic regression"
}
File diff suppressed because it is too large Load Diff
File diff suppressed because it is too large Load Diff
File diff suppressed because it is too large Load Diff
Binary file not shown.
@@ -0,0 +1,131 @@
"""Self-contained SkillBundle-v1 reader. Only Python stdlib and PyTorch required.
This file is copied verbatim into each immutable release. No project/Kit import.
SHA256 detects corruption; it is not a signature or a sandbox for untrusted code.
"""
import hashlib
import json
import sys
from pathlib import Path
import torch
from torch import nn
class Actor(nn.Module):
def __init__(self, hidden=128):
super().__init__()
self.register_buffer("mean", torch.zeros(15))
self.register_buffer("std", torch.ones(15))
self.net = nn.Sequential(
nn.Linear(15, hidden), nn.ReLU(), nn.Linear(hidden, hidden), nn.ReLU(), nn.Linear(hidden, 3)
)
def forward(self, state):
return self.net((state - self.mean) / self.std).tanh()
def file_hash(path):
return hashlib.sha256(Path(path).read_bytes()).hexdigest()
class FrozenPolicy:
"""Stateless deterministic leg-token-v1 -> float32 [B,8] normalized PD command.
Caller owns sensor sampling, last-applied action, goal validity and 100Hz clock.
The package binds exact static geometry/controller/units, not just tensor shapes.
"""
def __init__(self, directory, *, part_id, manifest_sha256, device="cpu"):
root = Path(directory).resolve()
manifest = json.loads((root / "manifest.json").read_text())
if manifest["format"] != "skillfactory.skillbundle.v1" or manifest["interface"] != "go2-fl-reach-leg-token-v1":
raise ValueError("Unsupported bundle/interface version")
if sys.version_info[:2] != (3, 12) or torch.__version__.split("+")[0] != "2.10.0":
raise RuntimeError("Validated dependency range: Python 3.12, torch 2.10.0")
for name, digest in manifest["files"].items():
path = (root / name).resolve()
if not path.is_relative_to(root) or not path.is_file() or file_hash(path) != digest:
raise ValueError(f"Bundle integrity failure: {name}")
required = {"actor.pt", "binding.json", "frozen_policy.py", "config.json", "README.md"}
if not required.issubset(manifest["files"]):
raise ValueError("Missing required integrity entries")
self.binding = json.loads((root / "binding.json").read_text())
if part_id != self.binding["part_id"] or manifest_sha256 != self.binding["manifest_sha256"]:
raise ValueError("Unsupported hardware binding")
self.device = torch.device(device)
self._actor = Actor(manifest["hidden"]).to(self.device)
self.device = next(self._actor.parameters()).device
self._actor.load_state_dict(
torch.load(root / "actor.pt", map_location=self.device, weights_only=True), strict=True
)
self._actor.eval().requires_grad_(False)
if any(not torch.isfinite(v).all() for v in self._actor.state_dict().values()) or (self._actor.std <= 0).any():
raise ValueError("Invalid weights or normalization state")
self._static = {
k: torch.tensor(v, device=self.device, dtype=torch.bool if "mask" in k else torch.float32)
for k, v in self.binding["static_tokens"].items()
}
self._context = torch.tensor(self.binding["context"], device=self.device, dtype=torch.float32)
self.version = manifest["version"]
@torch.inference_mode()
def act(self, tokens):
shapes = {
"joint_description": (8, 25),
"joint_observation": (8, 3),
"joint_mask": (8,),
"effector_description": (1, 16),
"effector_observation": (1, 11),
"effector_mask": (1,),
"context": (10,),
"goal": (3,),
}
if set(tokens) != set(shapes):
raise ValueError("Require all and only leg-token-v1 fields")
first = tokens["goal"]
if not isinstance(first, torch.Tensor) or first.ndim != 2 or len(first) < 1:
raise ValueError("Require nonempty batched tensors")
count = len(first)
for key, shape in shapes.items():
value = tokens[key]
dtype = torch.bool if "mask" in key else torch.float32
if (
not isinstance(value, torch.Tensor)
or value.shape != (count, *shape)
or value.dtype != dtype
or value.device != self.device
):
raise ValueError(f"Bad shape/dtype/device: {key}")
# Padded garbage is deliberately ignored; valid entities must be finite.
active = value[:, :3] if key.startswith("joint_") else value
if not torch.isfinite(active).all():
raise ValueError(f"Non-finite active input: {key}")
for key, expected in self._static.items():
actual = tokens[key]
if key == "joint_description":
actual, expected = actual[:, :3], expected[:3]
if not torch.equal(actual, expected.expand_as(actual)):
raise ValueError(f"Static binding mismatch: {key}")
if not torch.allclose(tokens["context"], self._context.expand_as(tokens["context"]), atol=1e-5, rtol=0):
raise ValueError("Context mismatch: fixed suspended anchor/control period/physical scales required")
x = torch.cat(
(tokens["joint_observation"][:, :3].flatten(1), tokens["effector_observation"][:, 0, :3], tokens["goal"]),
-1,
)
result = torch.zeros(count, 8, device=self.device, dtype=torch.float32)
result[:, :3] = self._actor(x)
if not torch.isfinite(result).all():
raise RuntimeError("Non-finite policy output")
return result
@torch.inference_mode()
def position_targets(self, tokens):
"""Native declared-order rad targets. Actual driver must still enforce torque/velocity limits."""
action = self.act(tokens)[:, :3].clamp(-1, 1)
controls = self.binding["spec"]["joints"]
nominal = action.new_tensor([j["q_default"] for j in controls])
scale = action.new_tensor([j["action_scale"] for j in controls])
soft = action.new_tensor([j["soft_limits"] for j in self.binding["joints"]])
return (nominal + scale * action).clamp(min=soft[:, 0], max=soft[:, 1])
+38
View File
@@ -0,0 +1,38 @@
{
"format": "skillfactory.skillbundle.v1",
"interface": "go2-fl-reach-leg-token-v1",
"version": "1.0.0",
"status": "accepted",
"frozen": true,
"hidden": 128,
"dependencies": {
"python": "3.12.x",
"torch": "2.10.0 (CPU validated; CUDA evaluated separately)"
},
"scope": "Go2 FL only; fixed suspended anchor; 3-DoF foot-link Cartesian stationary reach; nominal reset; no contact, noise, morphology or whole-body transfer claims",
"files": {
"ASSET_LICENSE": "84aac59fd3246e3ddc49d1387644e8fcf43b0def4f0b9fe687f372e90446df2d",
"README.md": "7c6cfaea271595f106e4f042afb94a4e16916c75405d17425a02fd08acf37dde",
"actor.pt": "cdfe459887e7629924df7e6a79fd2900eff184e5b25bc2a26bb585378f671c4a",
"binding.json": "a8b6decc1e311b853943e088b48773a2de78b7ed1d67e465f53b2d849f242dbd",
"config.json": "1c56a59a74cb35810b3a43e409509c03b9f96bcb285a2e09b887e01cef9df114",
"evidence/boundary.json": "978bc4359081e62496fb335ef2ae109ca35437f5044254a20f240f19ab2e9e64",
"evidence/data_dataset.json": "1228d9ee3b2cb65b6b98d9d193e7b66f8506a35436c6c5edea3c3816f5b60023",
"evidence/engineering_checks.json": "1d5bf66e9e382db26d69e5be23a96a8e96b2447fd0c10b8a5fcbce202a9e2013",
"evidence/inference_dependencies.txt": "71e6736e0da59e90e625b370c2ba5552183c13d3cf61327818d296a1ffce1d3a",
"evidence/provenance.json": "a8ad5ee503d64996c6077040cff1ffb1012466c1af4e66b458e0c275d4566c46",
"evidence/resume_exact.json": "339346a94987ee04d4ed0761eeb4a81aac0048351fffb1e8b2e21b970e3718fd",
"evidence/stability.json": "b350a501b2cb7381b7d7e060c38700f2f431c916fc57802b98f8364012273c9e",
"evidence/standalone.json": "ac8f980765567996fb95853661bbff5d559238f2b2ec7b69501848d953b0ea84",
"evidence/test.json": "b1c8cf1dd3a61800b13a971dde62d372c450c029ffbb2b4464175d67ff644264",
"evidence/tested_candidate_manifest.json": "ca84aa6c5b6e87cdd2238bc6848c9f046d6b20eb71e4c66ea090631b81cc172b",
"evidence/train_result.json": "5a110a7dfe373e288418f235f7db90909b578fa2da7443674247ea0e03886aee",
"evidence/validation_010000.json": "e1ba15034bf350e78ad1cb21c1fe906c15ddf6b03d373f661cf940941eb1eaba",
"evidence/validation_020000.json": "62898992738a4005d6a32e10303081911ffe0f1d5ac3b1bc7894b2d3c96cd7c7",
"evidence/validation_030000.json": "3d3943fdeea405578374edd016d182b1fe3b175f09b7c748ddbac54ca9ddf831",
"example_tokens.pt": "f91ac1f14f9dc63d9f857f5cf901573ef21a219f6618e506bd0b92b57df9538f",
"frozen_policy.py": "562d4ff6e4b5d1d569e16badcc44c972d303284e06371be22060a682aae38ffb",
"requirements.txt": "69191f795f024589b482e6bf90c5db6a5631c0296f383f193eb1c1f76dfeaefd",
"selection.json": "cdf8a22ac773b88322556ec77a35e5247814a13258bd684778754e50413ff4c6"
}
}
@@ -0,0 +1 @@
torch==2.10.0
@@ -0,0 +1,20 @@
{
"selected_update": 20000,
"checkpoint_sha256": "a1cb8076149bea872e6300c66984214828ceea686972f585970552bc52f90e9d",
"actor_sha256": "cdfe459887e7629924df7e6a79fd2900eff184e5b25bc2a26bb585378f671c4a",
"rule": "lowest final-window mean error among validation candidates satisfying all nominal gates; no selection on test",
"selection_seeds": [
21000,
21001
],
"config_sha256": "7e3e47a6d5b377dcdef98639dbb72cd585834c6e45b9750283f195bcb2cdcd52",
"training_identity": {
"config_sha256": "7e3e47a6d5b377dcdef98639dbb72cd585834c6e45b9750283f195bcb2cdcd52",
"dataset_sha256": "1228d9ee3b2cb65b6b98d9d193e7b66f8506a35436c6c5edea3c3816f5b60023",
"source_sha256": {
"scripts/train_offline_leg.py": "cf05cd55c6d5d875417928bc9d31d54d1d7c073d2877563d749bdf70e556a2b5",
"source/SkillFactory/SkillFactory/legs/offline.py": "12ac018a2c76d046c8ed50ee105fcd6a33931f7c7a3bb027b6db332726ab064d",
"source/SkillFactory/SkillFactory/legs/frozen_policy.py": "562d4ff6e4b5d1d569e16badcc44c972d303284e06371be22060a682aae38ffb"
}
}
}
File diff suppressed because it is too large Load Diff
@@ -0,0 +1,26 @@
{
"passed": true,
"actor_sha256": "cdfe459887e7629924df7e6a79fd2900eff184e5b25bc2a26bb585378f671c4a",
"manifest_sha256": "b685088e675fc624744a239644ce59766efab6bbb8f4790b4783c16f0ed100fc",
"python": "3.12.3 (main, Aug 31 2026, 10:18:26) [GCC 13.3.0]",
"torch": "2.10.0+cpu",
"device": "cpu",
"isolated": true,
"parameters_and_buffers_unchanged": true,
"requires_grad": false,
"training_or_simulator_imports": false,
"checks": [
"B=1:100 deterministic calls; no input mutation",
"B=16:100 deterministic calls; no input mutation",
"B=64:100 deterministic calls; no input mutation",
"B=256:100 deterministic calls; no input mutation",
"invalid active values/dtype/masks/context/hardware rejected",
"padding ignored",
"corruption rejected"
],
"package_only": true,
"run_id": "fc58460e959c4211bbf13be6d30c9119",
"package": "/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/releases/go2_fl_reach-1.0.1",
"stderr": "/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/stage2/inference_venv/lib/python3.12/site-packages/torch/_subclasses/functional_tensor.py:283: UserWarning: Failed to initialize NumPy: No module named 'numpy' (Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:84.)\n cpu = _conversion_method_template(device=torch.device(\"cpu\"))\n",
"status": "completed"
}
Binary file not shown.
+29
View File
@@ -0,0 +1,29 @@
BSD 3-Clause License
Copyright (c) 2016-2022 HangZhou YuShu TECHNOLOGY CO.,LTD. ("Unitree Robotics")
All rights reserved.
Redistribution and use in source and binary forms, with or without
modification, are permitted provided that the following conditions are met:
* Redistributions of source code must retain the above copyright notice, this
list of conditions and the following disclaimer.
* Redistributions in binary form must reproduce the above copyright notice,
this list of conditions and the following disclaimer in the documentation
and/or other materials provided with the distribution.
* Neither the name of the copyright holder nor the names of its
contributors may be used to endorse or promote products derived from
this software without specific prior written permission.
THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS "AS IS"
AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT LIMITED TO, THE
IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS FOR A PARTICULAR PURPOSE ARE
DISCLAIMED. IN NO EVENT SHALL THE COPYRIGHT HOLDER OR CONTRIBUTORS BE LIABLE
FOR ANY DIRECT, INDIRECT, INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL
DAMAGES (INCLUDING, BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR
SERVICES; LOSS OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER
CAUSED AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT OF THE USE
OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF SUCH DAMAGE.
+34
View File
@@ -0,0 +1,34 @@
# Go2 FL Reach SkillBundle 1.0.1
仅固定悬空 Go2 FL,100Hz控制/500Hz物理、名义复位、2秒定点目标。
目标为 anchor frame 足端 link xyz/L,由 q_nominal+0.25*a,a∈[-.75,.75]^3 的FK生成;不是关节目标输入。
不支持接触承重、Go1、其他形态、移动根、噪声、步态、整机或实机。无 latent 技能库声明。
依赖:Python3.12,torch==2.10.0(其传递依赖由 pip 安装);CPU无Isaac依赖。
`python -m pip install torch==2.10.0 --index-url https://download.pytorch.org/whl/cpu`
只把本包目录加入 sys.path,然后:
```python
from frozen_policy import FrozenPolicy
import json, torch
from pathlib import Path
root = Path('/path/to/go2_fl_reach-1.0.1')
binding = json.loads((root/'binding.json').read_text())
p = FrozenPolicy(root, part_id='go2_fl', manifest_sha256=binding['manifest_sha256'])
tokens = torch.load(root/'example_tokens.pt', weights_only=True) # 仅接口示例;实际须提供实时状态/目标
command = p.act(tokens) # [B,8] float32,后5槽0;前3槽声明顺序、[-1,1]
q_target = p.position_targets(tokens) # [B,3] rad,已按soft_limits裁剪
```
实际调用方必须使用**自身资产的**manifest hash(示例从包读取仅供离线演示),按binding.joint_order映射。
输入完整旧leg-token-v1字典,字段/特征顺序见binding.feature_schema;mask bool,其余float32,同device。
归一化为固定物理尺度再经训练集mean/std;全部必要状态保存在actor.pt,推理无统计/参数更新。
joint_observation为(q-q_nominal)/scale、dq/URDF限速、last_applied_action;goal与foot位置均除length_scale_m。
静态tokens/context须匹配绑定;有效槽NaN/Inf、错shape/dtype/版本/硬件hash被拒绝,padding垃圾忽略。
执行器为显式IdealPD,Kp/Kd/effort/velocity/soft limits见binding;不能把输出直接当电机力矩。
调用方维持控制时序、有效传感、动作历史与驱动限矩限速。无实机watchdog/故障安全认证。
这是独立SkillBundle v1,不是声称通用SoftwareSpec v2 adapter已经实现。
完整性:manifest.json列出所有有效负载SHA256;外部SHA256SUMS固定manifest的信任锚。
Hash防意外损坏,不是签名;只执行可信来源的Python代码。拒绝覆盖已有发布目录。
任何权重、归一化、接口或元数据变化应发布新版本。查验manifest.status,candidate不是验收发布。
Binary file not shown.
+499
View File
@@ -0,0 +1,499 @@
{
"part_id": "go2_fl",
"manifest_sha256": "4cea30b0264444c2e0417d0c26d6ea9148698d3b77f559f12a9b33775a41e3eb",
"joint_order": [
"FL_hip_joint",
"FL_thigh_joint",
"FL_calf_joint"
],
"spec": {
"auxiliary_root_joints": [
"FL_hip_rotor_joint"
],
"decimation": 5,
"foot_body": "FL_foot",
"foot_collision_index": 0,
"joints": [
{
"action_scale": 0.25,
"kd": 0.5,
"kp": 25.0,
"name": "FL_hip_joint",
"q_default": 0.0
},
{
"action_scale": 0.25,
"kd": 0.5,
"kp": 25.0,
"name": "FL_thigh_joint",
"q_default": 0.8
},
{
"action_scale": 0.25,
"kd": 0.5,
"kp": 25.0,
"name": "FL_calf_joint",
"q_default": -1.5
}
],
"license_file": "../../LICENSE",
"license_id": "BSD-3-Clause",
"mount_joint": "FL_hip_joint",
"package_name": "go2_description",
"part_id": "go2_fl",
"physics_dt": 0.002,
"schema_version": 1,
"soft_limit_factor": 0.9,
"source_root": "/home/cen/Embodied_Workspace/unitree_ros/robots/go2_description",
"urdf": "urdf/go2_description.urdf"
},
"joints": [
{
"axis": [
1.0,
0.0,
0.0
],
"child": "FL_hip",
"limits": {
"effort": 23.7,
"lower": -1.0472,
"upper": 1.0472,
"velocity": 30.1
},
"name": "FL_hip_joint",
"nominal_axis": [
1.0,
0.0,
0.0
],
"nominal_position": [
0.0,
0.0,
0.0
],
"soft_limits": [
-0.94248,
0.94248
]
},
{
"axis": [
0.0,
1.0,
0.0
],
"child": "FL_thigh",
"limits": {
"effort": 23.7,
"lower": -1.5708,
"upper": 3.4907,
"velocity": 30.1
},
"name": "FL_thigh_joint",
"nominal_axis": [
0.0,
1.0,
0.0
],
"nominal_position": [
0.0,
0.0955,
0.0
],
"soft_limits": [
-1.3177249999999998,
3.237625
]
},
{
"axis": [
0.0,
1.0,
0.0
],
"child": "FL_calf",
"limits": {
"effort": 45.43,
"lower": -2.7227,
"upper": -0.83776,
"velocity": 15.7
},
"name": "FL_calf_joint",
"nominal_axis": [
0.0,
1.0,
0.0
],
"nominal_position": [
-0.15279684736159835,
0.0955,
-0.14839852909094622
],
"soft_limits": [
-2.6284530000000004,
-0.9320069999999999
]
}
],
"length_scale_m": 0.5455,
"source": {
"commit": "daadf41ee9afce8f90fdc09a98506012691fa122",
"license_id": "BSD-3-Clause",
"license_sha256": "84aac59fd3246e3ddc49d1387644e8fcf43b0def4f0b9fe687f372e90446df2d",
"sha256": "7d19fe48e2e689ee1a032ab99f2a4a8b671d87e73de48d3e65811682a5b48b9e",
"urdf": "/home/cen/Embodied_Workspace/unitree_ros/robots/go2_description/urdf/go2_description.urdf"
},
"feature_schema": {
"version": 1,
"joint_description": [
"pivot_x/L",
"pivot_y/L",
"pivot_z/L",
"axis_x",
"axis_y",
"axis_z",
"q_nominal/pi",
"control_min/pi",
"control_max/pi",
"effort/Emax",
"velocity/Vmax",
"kp*pi/Emax",
"kd*Vmax/Emax",
"action_scale/pi",
"child_mass/Mknown",
"child_Ixx/ML2",
"child_Ixy/ML2",
"child_Ixz/ML2",
"child_Iyy/ML2",
"child_Iyz/ML2",
"child_Izz/ML2",
"child_com_x/L",
"child_com_y/L",
"child_com_z/L",
"inertial_valid"
],
"joint_observation": [
"(q-q_nominal)/action_scale",
"dq/velocity_limit",
"last_applied_action"
],
"effector_description": [
"nominal_x/L",
"nominal_y/L",
"nominal_z/L",
"collision_x/L",
"collision_y/L",
"collision_z/L",
"collision_qx",
"collision_qy",
"collision_qz",
"collision_qw",
"sphere_radius/L",
"box_x/L",
"box_y/L",
"box_z/L",
"is_sphere",
"is_box"
],
"effector_observation": [
"link_x/L",
"link_y/L",
"link_z/L",
"link_vx/(L/s)",
"link_vy/(L/s)",
"link_vz/(L/s)",
"normal_Fx/(Mknown*g)",
"normal_Fy/(Mknown*g)",
"normal_Fz/(Mknown*g)",
"contact",
"selected_collision_gap/L"
],
"context": [
"gravity_x",
"gravity_y",
"gravity_z",
"control_dt_s",
"ground_z_relative_to_anchor/L",
"length_scale_m",
"known_mass_scale_kg",
"max_effort/(Mknown*g*L)",
"max_joint_velocity_rad_per_s",
"has_missing_source_inertia"
],
"goal": [
"target_link_x/L",
"target_link_y/L",
"target_link_z/L"
],
"force_semantics": "normal contact force only; no friction wrench",
"inertia_semantics": "Source child COM inertia and relative COM offset rotated into anchor nominal frame; not fixed-subtree aggregated",
"mass_semantics": "Known source mass excludes synthetic anchor; not a prediction of support force"
},
"max_joints": 8,
"context": [
0.0,
0.0,
-1.0,
0.009999999776482582,
-1.183318018913269,
0.5454999804496765,
2.2909998893737793,
3.7055602073669434,
30.100000381469727,
1.0
],
"static_tokens": {
"joint_description": [
[
0.0,
0.0,
0.0,
1.0,
0.0,
0.0,
0.0,
-0.07957746833562851,
0.07957746833562851,
0.5216817259788513,
1.0,
1.7288094758987427,
0.33127889037132263,
0.07957746833562851,
0.295940637588501,
0.0007040874916128814,
-4.415215244080173e-06,
1.6282023125313572e-06,
0.0012966945068910718,
-2.0829254481213866e-06,
0.0008742419886402786,
-0.00989917479455471,
0.0035563702695071697,
-0.000192483959835954,
1.0
],
[
0.0,
0.17506875097751617,
0.0,
0.0,
1.0,
0.0,
0.2546479105949402,
0.17507043480873108,
0.3342253863811493,
0.5216817259788513,
1.0,
1.7288094758987427,
0.33127889037132263,
0.07957746833562851,
0.5028371810913086,
0.004511878360062838,
0.0009393356740474701,
-0.0035138893872499466,
0.008507723920047283,
0.0007339900475926697,
0.005565374158322811,
-0.04777860268950462,
-0.040879927575588226,
-0.03684582561254501,
1.0
],
[
-0.28010421991348267,
0.17506875097751617,
-0.27204129099845886,
0.0,
1.0,
0.0,
-0.47746482491493225,
-0.5570423007011414,
-0.39788734912872314,
1.0,
0.5215947031974792,
1.7288094758987427,
0.33127889037132263,
0.07957746833562851,
0.0672195553779602,
0.0009218948544003069,
-7.442903097398812e-06,
0.000761082861572504,
0.0016135338228195906,
9.610687811800744e-06,
0.0007105613476596773,
0.14349472522735596,
-0.0017873510951176286,
-0.15476909279823303,
1.0
],
[
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0
],
[
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0
],
[
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0
],
[
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0
],
[
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0,
0.0
]
],
"joint_mask": [
true,
true,
true,
false,
false,
false,
false,
false
],
"effector_description": [
[
-0.028558166697621346,
0.17506875097751617,
-0.5706872940063477,
-0.0036663610953837633,
0.0,
0.0,
0.0,
0.0,
0.0,
1.0,
0.04032997414469719,
0.0,
0.0,
0.0,
1.0,
0.0
]
],
"effector_mask": [
true
]
}
}
+79
View File
@@ -0,0 +1,79 @@
{
"experiment": "go2_fl_reach_iql_v1_attempt2",
"scope": "Go2 FL only; fixed suspended anchor; 3-DoF foot-link Cartesian stationary reach; nominal reset; no contact, noise, morphology or whole-body transfer claims",
"bundle": "assets/generated/legs/go2_fl",
"max_joints": 8,
"num_envs": 256,
"episode_steps": 200,
"settle_steps": 150,
"goal_action_range": 0.75,
"collection": {
"seed": 12000,
"shards": 20,
"noise_std_by_quartile": [
0.0,
0.03,
0.1,
0.25
],
"feedback_gain": 1.5,
"velocity_gain_s": 0.04,
"integral_gain_s_inv": 8.0
},
"training": {
"seed": 42,
"updates": 30000,
"batch_size": 512,
"hidden": 128,
"lr": 0.0003,
"discount": 0.99,
"expectile": 0.7,
"temperature": 3.0,
"max_weight": 100.0,
"target_tau": 0.005,
"checkpoint_every": 5000
},
"selection": {
"seed": 21000,
"batches": 2,
"candidate_updates": [
10000,
20000,
30000
],
"rule": "lowest final-window mean error among validation candidates satisfying all nominal gates; no selection on test"
},
"test": {
"seed": 31000,
"batches": 4,
"boundary_seed": 41000,
"boundary_batches": 1,
"stability_seed": 51000,
"stability_envs": 64,
"stability_seconds": 120
},
"acceptance": {
"mean_error_m": 0.008,
"p95_error_m": 0.015,
"success_radius_m": 0.015,
"success_fraction": 0.95,
"baseline_error_reduction": 0.6,
"boundary_mean_error_m": 0.012,
"boundary_p95_error_m": 0.02,
"max_normalized_torque": 1.00001,
"max_normalized_velocity": 1.00001,
"joint_limit_tolerance_rad": 0.01,
"anchor_pose_max_error": 0.0001,
"invalid_states": 0,
"unexpected_terminations": 0,
"memory_growth_fraction": 0.05
},
"budget": {
"gpu_count": 1,
"gpu": "RTX 5080 16 GiB",
"wall_seconds": 5400,
"disk_gb": 10,
"max_training_attempts": 2
},
"release_version": "1.0.0"
}
File diff suppressed because it is too large Load Diff
@@ -0,0 +1,215 @@
{
"config_sha256": "7e3e47a6d5b377dcdef98639dbb72cd585834c6e45b9750283f195bcb2cdcd52",
"asset_manifest_sha256": "4cea30b0264444c2e0417d0c26d6ea9148698d3b77f559f12a9b33775a41e3eb",
"feature_schema": {
"version": 1,
"joint_description": [
"pivot_x/L",
"pivot_y/L",
"pivot_z/L",
"axis_x",
"axis_y",
"axis_z",
"q_nominal/pi",
"control_min/pi",
"control_max/pi",
"effort/Emax",
"velocity/Vmax",
"kp*pi/Emax",
"kd*Vmax/Emax",
"action_scale/pi",
"child_mass/Mknown",
"child_Ixx/ML2",
"child_Ixy/ML2",
"child_Ixz/ML2",
"child_Iyy/ML2",
"child_Iyz/ML2",
"child_Izz/ML2",
"child_com_x/L",
"child_com_y/L",
"child_com_z/L",
"inertial_valid"
],
"joint_observation": [
"(q-q_nominal)/action_scale",
"dq/velocity_limit",
"last_applied_action"
],
"effector_description": [
"nominal_x/L",
"nominal_y/L",
"nominal_z/L",
"collision_x/L",
"collision_y/L",
"collision_z/L",
"collision_qx",
"collision_qy",
"collision_qz",
"collision_qw",
"sphere_radius/L",
"box_x/L",
"box_y/L",
"box_z/L",
"is_sphere",
"is_box"
],
"effector_observation": [
"link_x/L",
"link_y/L",
"link_z/L",
"link_vx/(L/s)",
"link_vy/(L/s)",
"link_vz/(L/s)",
"normal_Fx/(Mknown*g)",
"normal_Fy/(Mknown*g)",
"normal_Fz/(Mknown*g)",
"contact",
"selected_collision_gap/L"
],
"context": [
"gravity_x",
"gravity_y",
"gravity_z",
"control_dt_s",
"ground_z_relative_to_anchor/L",
"length_scale_m",
"known_mass_scale_kg",
"max_effort/(Mknown*g*L)",
"max_joint_velocity_rad_per_s",
"has_missing_source_inertia"
],
"goal": [
"target_link_x/L",
"target_link_y/L",
"target_link_z/L"
],
"force_semantics": "normal contact force only; no friction wrench",
"inertia_semantics": "Source child COM inertia and relative COM offset rotated into anchor nominal frame; not fixed-subtree aggregated",
"mass_semantics": "Known source mass excludes synthetic anchor; not a prediction of support force"
},
"shards": [
{
"path": "shard_000.pt",
"sha256": "6c22000bd871a41b88b90851002db418d025988f15c02b77a83305c3106ed9b7",
"seed": 12000,
"transitions": 51200
},
{
"path": "shard_001.pt",
"sha256": "53ae0894c392db2823eeac22551a951fe0f93418b0d8a8674803d78855c1d2d5",
"seed": 12001,
"transitions": 51200
},
{
"path": "shard_002.pt",
"sha256": "69103fc3729308af87096a75539e143d5f2ddba09ce250048a9dec6829b12775",
"seed": 12002,
"transitions": 51200
},
{
"path": "shard_003.pt",
"sha256": "a48b0354dfb69ebe2f18539c652f3137f5cec1f2aa9208aba7ca6fd60cefebbe",
"seed": 12003,
"transitions": 51200
},
{
"path": "shard_004.pt",
"sha256": "1025160d3b3de8025e36316e5876af8ea4666123fd10037848b4c14a7fe68acd",
"seed": 12004,
"transitions": 51200
},
{
"path": "shard_005.pt",
"sha256": "21a32b080efbb96e30f1da487ce596a19f2bb62dbe712bbf1ede0aba5252cd31",
"seed": 12005,
"transitions": 51200
},
{
"path": "shard_006.pt",
"sha256": "d55dee7a9b02b47c017c7f23cc340c417972f35722c3e4d4e6dc40d63aa7c42e",
"seed": 12006,
"transitions": 51200
},
{
"path": "shard_007.pt",
"sha256": "0eb5026af22bd86c78a9cd4ff079a450ceaeef525bf467f45d089c6891ff8885",
"seed": 12007,
"transitions": 51200
},
{
"path": "shard_008.pt",
"sha256": "8cde8a3f84d4b9ef68f055382badc18b91a480d223be62abbb49b92e1527ded4",
"seed": 12008,
"transitions": 51200
},
{
"path": "shard_009.pt",
"sha256": "3ff78a3d10dcfc4b2b98d156236251fcfd1fe79603923f08767406d623fa3911",
"seed": 12009,
"transitions": 51200
},
{
"path": "shard_010.pt",
"sha256": "8206486abff6fa604fd3f4c18ad107d00faa58bdcbd1269b38a5cbbdb3165517",
"seed": 12010,
"transitions": 51200
},
{
"path": "shard_011.pt",
"sha256": "c001d4561c129f08ed3dc62ed118e4bc9169efd0e5990a27b48225f4874fddda",
"seed": 12011,
"transitions": 51200
},
{
"path": "shard_012.pt",
"sha256": "f4c25b40b84b47d759e74f4b57f84772624431707fb1bb3da36f41606c7661a4",
"seed": 12012,
"transitions": 51200
},
{
"path": "shard_013.pt",
"sha256": "d7f4513cffe1fdcd3ffd2b4c7a9f2e32a204184e1256c3992f3aed68d6effa97",
"seed": 12013,
"transitions": 51200
},
{
"path": "shard_014.pt",
"sha256": "5f80cc08f31572ba74225f9acf8b80c17aee3d3347e90088cad390baa9a270a9",
"seed": 12014,
"transitions": 51200
},
{
"path": "shard_015.pt",
"sha256": "faf2523f065132f97e9e69c746da6404b0790c78cdf8d5aec9612e32d0e34a56",
"seed": 12015,
"transitions": 51200
},
{
"path": "shard_016.pt",
"sha256": "0a95607f01d6740fdcc2957757860b99f9a9ebdb47077918c9888f4001a7712e",
"seed": 12016,
"transitions": 51200
},
{
"path": "shard_017.pt",
"sha256": "68780da8bd6541343f71e35e6d4ff5525859a61e52a4456f1165848a3f099894",
"seed": 12017,
"transitions": 51200
},
{
"path": "shard_018.pt",
"sha256": "4d329c699320fca3432f6f4510a08dbff85c5311922fae901446611c39876b46",
"seed": 12018,
"transitions": 51200
},
{
"path": "shard_019.pt",
"sha256": "539b3a8c311878cffcffe534f7ace2bf12a6361013b7fc7c0f231b64313133d7",
"seed": 12019,
"transitions": 51200
}
],
"complete": true,
"transitions": 1024000,
"bootstrap": "No true terminations. Artificial 2s truncations bootstrap from pre-reset next-state."
}
@@ -0,0 +1,10 @@
{
"passed": true,
"pytest": "172 passed, 3 skipped, 15 warnings",
"pytest_log_sha256": "b17bd316537ba859d957d2e17f9c70029d970306d4e4b85219b17ba19052dfcc",
"ruff_log_sha256": "a5b097bbb606a1132331f167e0dab21849bbe45b37cc1957f17dcc665ece6218",
"git_diff_check": true,
"review_blockers_remaining": 0,
"review_kind": "local code inspection and regression tests; no separate external reviewer this turn",
"historical_optional_kit_tests_skipped": 3
}
@@ -0,0 +1,10 @@
filelock==3.32.3
fsspec==2026.7.0
Jinja2==3.1.6
MarkupSafe==3.0.3
mpmath==1.3.0
networkx==3.6.1
setuptools==78.1.0
sympy==1.14.0
torch==2.10.0+cpu
typing_extensions==4.16.0
@@ -0,0 +1,83 @@
{
"version": "1.0.1",
"purpose": "complete boundary revalidation and reliability-gated republication of unchanged 20k actor",
"training_origin": "logs/stage2/v1_attempt2",
"training_rerun": false,
"training_updates": 30000,
"selected_update": 20000,
"selection_rerun": false,
"export_selection": "reconfirmed historical validation-only rule; no selection from new test results",
"same_holdout_seeds": true,
"unseen_test_claim": false,
"training_config_sha256": "7e3e47a6d5b377dcdef98639dbb72cd585834c6e45b9750283f195bcb2cdcd52",
"source_hashes": {
"scripts/stage2_sim.py": "69a1ef05779e5869502d2d693dd7d2c5cc84bb36f0acbf0b918f96aa44b37373",
"scripts/stage2_boundary.py": "3c8c65581c82bdf689a2d8fe590f324acef5629e5e218675664218751815c326",
"scripts/stage2_acceptance.py": "e716602abfc410713bb1f6f79bb8958433f0ed4ea40e237c3d55b305d2b41b53",
"scripts/export_leg_bundle.py": "c5b0209e4770661098a626b0efa4d5055807098aaea7d9d4f2e57575a87e79ee",
"scripts/verify_frozen_leg.py": "f9f3df78646dca7a8117bed74259f2300423edbdea0f2905bebe41f7dbe80d16",
"scripts/verify_stage2_delivery.py": "0a35c13d0b789a62ea6c3012526bf8261963cc97d44589610828544122a26577",
"scripts/runtime_common.py": "57934860bc6fe1df64128376737aa2a80aafe8f85cf4bee38d7f4373edfe908b",
"tests/test_stage2_release.py": "80ca0d96ea4abbc05789ddbb312012fa19147c551583bf3d0c5a6e5b0e3cefee",
"tests/test_offline_leg.py": "94d4290866bcd7bf651a330eefb90c007ebfb5047227b86ca79ae9809f0064b0",
"scripts/train_offline_leg.py": "cf05cd55c6d5d875417928bc9d31d54d1d7c073d2877563d749bdf70e556a2b5",
"source/SkillFactory/SkillFactory/legs/offline.py": "12ac018a2c76d046c8ed50ee105fcd6a33931f7c7a3bb027b6db332726ab064d",
"source/SkillFactory/SkillFactory/legs/frozen_policy.py": "562d4ff6e4b5d1d569e16badcc44c972d303284e06371be22060a682aae38ffb"
},
"inherited_evidence_hashes": {
"data/shard_004.pt": "1025160d3b3de8025e36316e5876af8ea4666123fd10037848b4c14a7fe68acd",
"data/shard_018.pt": "4d329c699320fca3432f6f4510a08dbff85c5311922fae901446611c39876b46",
"data/shard_001.pt": "53ae0894c392db2823eeac22551a951fe0f93418b0d8a8674803d78855c1d2d5",
"data/shard_007.pt": "0eb5026af22bd86c78a9cd4ff079a450ceaeef525bf467f45d089c6891ff8885",
"data/shard_014.pt": "5f80cc08f31572ba74225f9acf8b80c17aee3d3347e90088cad390baa9a270a9",
"data/binding.json": "a8b6decc1e311b853943e088b48773a2de78b7ed1d67e465f53b2d849f242dbd",
"data/shard_016.pt": "0a95607f01d6740fdcc2957757860b99f9a9ebdb47077918c9888f4001a7712e",
"data/shard_017.pt": "68780da8bd6541343f71e35e6d4ff5525859a61e52a4456f1165848a3f099894",
"data/shard_015.pt": "faf2523f065132f97e9e69c746da6404b0790c78cdf8d5aec9612e32d0e34a56",
"data/result.json": "7cbfcd8f771204161f582225bbda0f42cdd3f2aa95ab4c3135148ac9a4e1da37",
"data/shard_002.pt": "69103fc3729308af87096a75539e143d5f2ddba09ce250048a9dec6829b12775",
"data/example_tokens.pt": "f91ac1f14f9dc63d9f857f5cf901573ef21a219f6618e506bd0b92b57df9538f",
"data/shard_019.pt": "539b3a8c311878cffcffe534f7ace2bf12a6361013b7fc7c0f231b64313133d7",
"data/shard_010.pt": "8206486abff6fa604fd3f4c18ad107d00faa58bdcbd1269b38a5cbbdb3165517",
"data/shard_011.pt": "c001d4561c129f08ed3dc62ed118e4bc9169efd0e5990a27b48225f4874fddda",
"data/shard_005.pt": "21a32b080efbb96e30f1da487ce596a19f2bb62dbe712bbf1ede0aba5252cd31",
"data/dataset.json": "1228d9ee3b2cb65b6b98d9d193e7b66f8506a35436c6c5edea3c3816f5b60023",
"data/shard_008.pt": "8cde8a3f84d4b9ef68f055382badc18b91a480d223be62abbb49b92e1527ded4",
"data/shard_009.pt": "3ff78a3d10dcfc4b2b98d156236251fcfd1fe79603923f08767406d623fa3911",
"data/shard_013.pt": "d7f4513cffe1fdcd3ffd2b4c7a9f2e32a204184e1256c3992f3aed68d6effa97",
"data/shard_012.pt": "f4c25b40b84b47d759e74f4b57f84772624431707fb1bb3da36f41606c7661a4",
"data/shard_006.pt": "d55dee7a9b02b47c017c7f23cc340c417972f35722c3e4d4e6dc40d63aa7c42e",
"data/shard_003.pt": "a48b0354dfb69ebe2f18539c652f3137f5cec1f2aa9208aba7ca6fd60cefebbe",
"data/shard_000.pt": "6c22000bd871a41b88b90851002db418d025988f15c02b77a83305c3106ed9b7",
"train/step_020000.pt": "a1cb8076149bea872e6300c66984214828ceea686972f585970552bc52f90e9d",
"train/step_005000.pt": "92c3fa0e445495d5fecd2f853d944aaf5b665f2dfef31d55334ec61b572aa623",
"train/result.json": "5a110a7dfe373e288418f235f7db90909b578fa2da7443674247ea0e03886aee",
"train/step_025000.pt": "812f18ef35786d8ef93f0d256c80feb5b81c65277462fe71b1bb2dff4816ea0a",
"train/step_030000.pt": "585e8ff49d1d51c3bd32743dbea82aaf8c14a4aab00943e0ba200a4375d8289f",
"train/latest.pt": "5dce2d6a09e58a96aae13f2a63747de3f42bb79ace76ae8859b6d5e8dd674746",
"train/events.jsonl": "587f6bf0c6bc9d4d5aae1065d4008954d5ca2ff4af3e182f41ffee5ccc6f2b75",
"train/step_015000.pt": "4a5dcbdb4f69ffd7d8ce398d2943bc0d315e653c189f642563b99270fce273d2",
"train/step_010000.pt": "45dc9e11738e1b0b8a56ceab5f7d49295e009ef06676fa34a00cc9a7eaf6cfe1",
"validation_010000/result.json": "e1ba15034bf350e78ad1cb21c1fe906c15ddf6b03d373f661cf940941eb1eaba",
"validation_020000/result.json": "62898992738a4005d6a32e10303081911ffe0f1d5ac3b1bc7894b2d3c96cd7c7",
"validation_030000/result.json": "3d3943fdeea405578374edd016d182b1fe3b175f09b7c748ddbac54ca9ddf831"
},
"fresh_evidence": [
"boundary",
"test",
"stability",
"standalone.json",
"resume_exact.json",
"cpu_cuda_parity.json",
"engineering_checks.json"
],
"source_base_commit": "8019d6f8d6b4004ab09137b35faa87a3117f2b56",
"worktree_dirty": true,
"no_git_commit_created": true,
"operational_scripts": {
"run_physics.py": "f61b4ae2878fb92fc7893c841abec9635873e38fef5a5a1c5ed1b27a9025e6d6",
"verify_reuse.py": "6689232d008c08e65bf1d7dcfc553098e69153273491d78efffa2cbb95163f23"
},
"runtime_cleanup": "own adopted helpers terminated/reaped; process reports record before/after; not a claim of leak-free Kit shutdown",
"old_release_manifest_sha256": "3106158d2ab214b09360528fb3c994a713435f05f801a8bde5115ebd16079655"
}
@@ -0,0 +1,24 @@
{
"passed": true,
"status": "completed",
"source_update": 5000,
"compared_update": 5001,
"full_learner_optimizer_rng_exact": true,
"tensor_comparisons": 112,
"device": "cuda:0",
"source_checkpoint_sha256": "92c3fa0e445495d5fecd2f853d944aaf5b665f2dfef31d55334ec61b572aa623",
"probe_checkpoint_sha256": {
"a": "38a611bd74ddc39218fa92a75776c740a8e8134663a5c46c28e2ef78bf574fbf",
"b": "0f23ec06a35ab7c15c37be73b815f7bbd92eeec57e838ded4c09d392414e6a2a"
},
"identity": {
"config_sha256": "7e3e47a6d5b377dcdef98639dbb72cd585834c6e45b9750283f195bcb2cdcd52",
"dataset_sha256": "1228d9ee3b2cb65b6b98d9d193e7b66f8506a35436c6c5edea3c3816f5b60023",
"source_sha256": {
"scripts/train_offline_leg.py": "cf05cd55c6d5d875417928bc9d31d54d1d7c073d2877563d749bdf70e556a2b5",
"source/SkillFactory/SkillFactory/legs/offline.py": "12ac018a2c76d046c8ed50ee105fcd6a33931f7c7a3bb027b6db332726ab064d",
"source/SkillFactory/SkillFactory/legs/frozen_policy.py": "562d4ff6e4b5d1d569e16badcc44c972d303284e06371be22060a682aae38ffb"
}
},
"note": "Two fresh processes; probes are not new training or policy selection."
}
@@ -0,0 +1,23 @@
{
"max_normalized_torque": 0.6637091636657715,
"max_normalized_velocity": 1.0000008344650269,
"joint_limit_excess_rad": 0.0,
"anchor_pose_error": 0.0,
"invalid_states": 0,
"physical_samples": 61000,
"control_steps": 12000,
"environment_control_steps": 768000,
"seconds_per_env": 120.0,
"num_envs": 64,
"device_used_before": 3525246976,
"device_used_max": 3554869248,
"memory_growth_fraction": 0.00840289267721366,
"weight_sha256": "cdfe459887e7629924df7e6a79fd2900eff184e5b25bc2a26bb585378f671c4a",
"package_manifest_sha256": "11396f964db7d8e26f9ca77b63dedb09a03ebe54ef3ebca11b0c4842b53617ce",
"passed": true,
"config_sha256": "7e3e47a6d5b377dcdef98639dbb72cd585834c6e45b9750283f195bcb2cdcd52",
"device": "cuda:0",
"asset_manifest_sha256": "4cea30b0264444c2e0417d0c26d6ea9148698d3b77f559f12a9b33775a41e3eb",
"wall_time_s": 170.5471037339994,
"status": "completed"
}
@@ -0,0 +1,26 @@
{
"passed": true,
"actor_sha256": "cdfe459887e7629924df7e6a79fd2900eff184e5b25bc2a26bb585378f671c4a",
"manifest_sha256": "11396f964db7d8e26f9ca77b63dedb09a03ebe54ef3ebca11b0c4842b53617ce",
"python": "3.12.3 (main, Aug 31 2026, 10:18:26) [GCC 13.3.0]",
"torch": "2.10.0+cpu",
"device": "cpu",
"isolated": true,
"parameters_and_buffers_unchanged": true,
"requires_grad": false,
"training_or_simulator_imports": false,
"checks": [
"B=1:100 deterministic calls; no input mutation",
"B=16:100 deterministic calls; no input mutation",
"B=64:100 deterministic calls; no input mutation",
"B=256:100 deterministic calls; no input mutation",
"invalid active values/dtype/masks/context/hardware rejected",
"padding ignored",
"corruption rejected"
],
"package_only": true,
"run_id": "fca4d22a022b421e8e8e6c7aa2178a38",
"package": "/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/stage2/v1_0_1/candidate",
"stderr": "/home/cen/Embodied_Workspace/Isaac_Projects/SkillFactory/logs/stage2/inference_venv/lib/python3.12/site-packages/torch/_subclasses/functional_tensor.py:283: UserWarning: Failed to initialize NumPy: No module named 'numpy' (Triggered internally at /pytorch/torch/csrc/utils/tensor_numpy.cpp:84.)\n cpu = _conversion_method_template(device=torch.device(\"cpu\"))\n",
"status": "completed"
}

Some files were not shown because too many files have changed in this diff Show More