246b48b1fb
Bring in kinematic hard-write control with mesh-convex collision projection, SDK mapping aligned to Isaac, launch/run scripts, and implementation/tech writeups.
500 lines
18 KiB
Markdown
500 lines
18 KiB
Markdown
# G20 左手 · Isaac Sim 技术详解
|
||
|
||
|
||
资产路径、URDF→USD 改动清单见 [`G20_Isaac_实现说明.md`](./G20_Isaac_实现说明.md)。
|
||
MuJoCo 对照:[`G20_MuJoCo_技术详解.md`](./G20_MuJoCo_技术详解.md)。
|
||
|
||
主文件:
|
||
|
||
| 角色 | 路径 |
|
||
|------|------|
|
||
| 桥主程序 | `/home/lxqs/isaacsim/linkerhand/isaac_g20_plan_bridge.py` |
|
||
| 关节映射 | `/home/lxqs/isaacsim/linkerhand/joint_mapping.py` |
|
||
| 启动脚本 | `/home/lxqs/isaacsim/linkerhand/run_g20_plan.sh` |
|
||
| 默认 USD | `.../urdf/g20/left/linkerhand_g20_left/linkerhand_g20_left.usda` |
|
||
|
||
---
|
||
|
||
## 1. 总体结构:单线程循环(与 MuJoCo 不同)
|
||
|
||
Isaac 桥**没有**单独的 ROS daemon 线程,也**没有** `threading.Lock`。
|
||
主循环里每个物理步先 `rclpy.spin_once`,再写关节、再 `SimulationManager.step()`。
|
||
|
||
```text
|
||
┌─────────────────────────────────────────────────────────────┐
|
||
│ 模块加载期(import 顶层) │
|
||
│ parse_args() │
|
||
│ SimulationApp(...) # 必须最早启动 Isaac │
|
||
│ enable_extension(ros2.bridge) │
|
||
│ import joint_mapping / rclpy │
|
||
├─────────────────────────────────────────────────────────────┤
|
||
│ main() 初始化 │
|
||
│ create_new_stage / GroundPlane / DistantLight │
|
||
│ _load_g20(usd) → Articulation │
|
||
│ _fix_base_to_world / set_dof_gains / 自碰 │
|
||
│ G20PlanBridge(...) │
|
||
│ apply_targets() 初始 snap 到 open │
|
||
├─────────────────────────────────────────────────────────────┤
|
||
│ 主循环(每 render 帧内跑 phys_per_render 次物理步) │
|
||
│ for _ in range(phys_per_render): │
|
||
│ rclpy.spin_once(node) # 处理回调 → 改 goal │
|
||
│ bridge.apply_targets() # slew → 写 Articulation │
|
||
│ SimulationManager.step() # PhysX 一步 │
|
||
│ recover_if_invalid() │
|
||
│ bridge.publish_state() │
|
||
│ if 位置硬写: apply_targets() 再钉一次 │
|
||
│ RenderingManager.render() │
|
||
│ simulation_app.update() │
|
||
└─────────────────────────────────────────────────────────────┘
|
||
```
|
||
|
||
**核心状态变量(`G20PlanBridge`):**
|
||
|
||
| 变量 | 类型 | 含义 |
|
||
|------|------|------|
|
||
| `goal` | `dict[str, float]` | 最新目标关节角(弧度,含 mimic) |
|
||
| `applied` | `dict[str, float]` | 本步实际写入的角(slew 后) |
|
||
| `targets` | 同 `applied` | 别名,历史兼容 |
|
||
| `track_nominal` | `dict[str, float] \| None` | 主动关节目标,供 `tracking_error()` |
|
||
| `dof_index` | `dict[str, int]` | 关节名 → Articulation DOF 下标 |
|
||
| `use_physics_drive` | `bool` | True=`--pd-drive`;False=位置硬写 |
|
||
|
||
**与 MuJoCo 的关键差异:**
|
||
|
||
| 项 | Isaac | MuJoCo |
|
||
|----|-------|--------|
|
||
| 线程模型 | 单线程 `spin_once` + 物理循环 | ROS daemon + 主仿真线程 + Lock |
|
||
| 防穿模 | USD 选择性 collider + 自碰 | `collision_project` + `mj_forward` |
|
||
| slew 单位 | **rad / 物理步**,默认 **0**(瞬达) | **rad / 秒**,默认 **5.0** |
|
||
| 写关节 API | `set_dof_positions` / `set_dof_position_targets` | `data.qpos` + `mj_forward` |
|
||
| 物理步进 | 每 render 默认 16 次 `SimulationManager.step` | 硬写模式不 `mj_step` |
|
||
|
||
---
|
||
|
||
## 2. 程序入口与加载顺序(很重要)
|
||
|
||
入口:`./run_g20_plan.sh` → `$ISAAC_SIM_DIR/python.sh isaac_g20_plan_bridge.py`。
|
||
|
||
### 2.1 为何 `parse_args` + `SimulationApp` 在顶层?
|
||
|
||
```text
|
||
# 文件顶层执行顺序(main 之前)
|
||
parse_args() # ① 先解析,因为 SimulationApp 要 headless
|
||
SimulationApp({"headless": ...}) # ② 启动 Omniverse / PhysX 运行时
|
||
... import isaacsim.* ...
|
||
from joint_mapping import ...
|
||
enable_extension("isaacsim.ros2.bridge")
|
||
simulation_app.update()
|
||
import rclpy ...
|
||
```
|
||
|
||
Isaac 要求:**先有 `SimulationApp`,才能 import 多数 isaacsim 模块**。
|
||
因此参数解析不能放在 `main()` 里才做;顶层 `args = parse_args()` 是刻意的。
|
||
|
||
`_normalize_argv`:把 `-pd-drive` 这类单横杠改成 `--pd-drive`,避免被 argparse 拆成 `-p -d ...`。
|
||
|
||
### 2.2 `main()` 调用链
|
||
|
||
```text
|
||
main()
|
||
├── stage_utils.create_new_stage()
|
||
├── GroundPlane / DistantLight
|
||
├── _compare_camera_eye_target() → set_camera_view(...)
|
||
├── hand_path = _load_g20(args.g20_usd)
|
||
│ ├── define_prim("/World/G20Hand")
|
||
│ ├── XformPrim(..., positions=[[0, 0.45, 0.70]])
|
||
│ ├── add_reference_to_stage(usd → .../hand)
|
||
│ ├── set_prim_variants(Physics=physx)
|
||
│ └── _patch_joint_max_forces(hand_path)
|
||
├── SimulationManager.set_physics_dt(1/120)
|
||
├── play → update → _fix_base_to_world → stop → play
|
||
├── art = Articulation(hand_path)
|
||
├── art.set_enabled_self_collisions(...)
|
||
├── art.set_link_enabled_gravities(...)
|
||
├── set_dof_gains / set_dof_max_efforts # 见 §5
|
||
├── rclpy.init(); Node("isaac_g20_plan_bridge")
|
||
├── G20PlanBridge(...)
|
||
├── slew=0; apply_targets(); 恢复 slew # 初始 snap open
|
||
└── while simulation_app.is_running():
|
||
for _ in range(phys_per_render): # 默认 16
|
||
spin_once → apply_targets → step → publish → [再钉]
|
||
render + update
|
||
```
|
||
|
||
---
|
||
|
||
## 3. 场景加载相关函数
|
||
|
||
### 3.1 `_load_g20(usd_path) → hand_path`
|
||
|
||
```python
|
||
mount = "/World/G20Hand"
|
||
stage_utils.define_prim(mount, "Xform")
|
||
XformPrim(mount, positions=[[0.0, HAND_MOUNT_Y, HAND_MOUNT_Z]], ...)
|
||
hand_path = f"{mount}/hand"
|
||
stage_utils.add_reference_to_stage(usd_path=..., path=hand_path)
|
||
prim_utils.set_prim_variants(hand_path, variants=[("Physics", "physx")])
|
||
_patch_joint_max_forces(hand_path)
|
||
return hand_path
|
||
```
|
||
|
||
要点:
|
||
|
||
1. **运行时不读 URDF**,只挂预烘焙 USD。
|
||
2. 手装在世界坐标 `(0, 0.45, 0.70)`,离地,避免和地面穿插。
|
||
3. Physics variant 强制 `physx`(USD 里还有 mujoco/newton 等层)。
|
||
|
||
### 3.2 `_patch_joint_max_forces(hand_path, max_force=8.0)`
|
||
|
||
遍历手下面所有 `UsdPhysics.RevoluteJoint`,取 angular DriveAPI:
|
||
|
||
```python
|
||
if maxForce < 8.0:
|
||
attr.Set(8.0)
|
||
```
|
||
|
||
防止某些关节 drive 力矩上限太小,PD 跟不上。
|
||
|
||
### 3.3 `_fix_base_to_world(hand_path)`
|
||
|
||
钉住掌心,手不会掉下去:
|
||
|
||
1. 在手 prim 树里找 `hand_base_link`(或 `base_link` / `palm`)RigidBody。
|
||
2. 若已有合法 `root_joint`:改 `LocalPos0/Rot0` 对齐当前世界位姿,Body1=base,Body0=世界。
|
||
3. 否则新建 `{hand_path}/WorldFixedJoint`(`UsdPhysics.FixedJoint`)。
|
||
|
||
调用时机:`play → update → fix → stop → play`,让 PhysX 在重建 articulation 前把固定关节写进 stage。
|
||
|
||
### 3.4 `_compare_camera_eye_target()`
|
||
|
||
与 MuJoCo `COMPARE_*` 对齐:
|
||
|
||
```python
|
||
target = hand_origin + COMPARE_LOOKAT_LOCAL # (0.04,-0.04,0.12) 相对手
|
||
eye.x = target.x + distance * cos(elev) # 从 +X 看
|
||
eye.z = target.z - distance * sin(elev) # elev=-12° 略俯视
|
||
```
|
||
|
||
`set_camera_view(eye, target, "/OmniverseKit_Persp")`;play/stop 后会再设一次,防止 viewport 被重置。
|
||
|
||
---
|
||
|
||
## 4. 映射层:`joint_mapping.py` 函数说明
|
||
|
||
与 MuJoCo `g20_joints.py` **数值一致**,但 API 多了 `hand`/`side` 参数(同一文件还服务 O6/L20)。
|
||
|
||
### 4.1 Legacy 0~255 调用链
|
||
|
||
```text
|
||
_on_cmd_u8(msg)
|
||
vals = [255]*20
|
||
优先按 msg.name → L20_SDK_TO_URDF 反查 SDK 索引填 vals
|
||
否则按下标填 vals[:20]
|
||
sdk_range_to_full_urdf_positions(vals, "G20", "left")
|
||
├── sdk_range_to_urdf_positions(...)
|
||
│ ├── range_to_arc(vals, "G20", "left")
|
||
│ └── SDK15: thumb_mcp = arc/1.02
|
||
│ SDK16..19: *_pip = arc/0.89
|
||
└── 补全 dip/ip(函数内部直接写,不等价于再调 complete_*,结果相同)
|
||
set_nominal_targets(full)
|
||
└── goal = complete_g20_mimic_positions(positions)
|
||
```
|
||
|
||
### 4.2 `range_to_arc(values, hand, side)`
|
||
|
||
```python
|
||
j_min, j_max, j_dir = _tables("G20", "left") # → G20_L_MIN/MAX/DIR
|
||
for i in 0..19:
|
||
if 11 <= i <= 14: continue
|
||
if j_dir[i] == -1:
|
||
arcs[i] = lerp(val, 0, 255, j_max[i], j_min[i])
|
||
else:
|
||
arcs[i] = lerp(val, 0, 255, j_min[i], j_max[i]) # 侧摆 6..9
|
||
```
|
||
|
||
### 4.3 `sdk_range_to_urdf_positions` vs `sdk_range_to_full_urdf_positions`
|
||
|
||
| 函数 | 输出 |
|
||
|------|------|
|
||
| `sdk_range_to_urdf_positions` | 仅主动 16 关节(已做 SDK15/16~19 除 mimic) |
|
||
| `sdk_range_to_full_urdf_positions` | 主动 + 软件解析的 `thumb_ip` / `*_dip` |
|
||
|
||
Isaac **硬写模式**写全关节(含 mimic);**PD 模式**只写主动关节,mimic 交给 PhysX USD mimic。
|
||
|
||
### 4.4 `complete_g20_mimic_positions`
|
||
|
||
与 MuJoCo 相同:`thumb_ip = thumb_mcp * 1.02`,`*_dip = *_pip * 0.89`;已有 mimic 值不覆盖。
|
||
|
||
### 4.5 `open_hand_command("G20")`
|
||
|
||
返回 `G20_OPEN_CMD`(float 列表),与 MuJoCo `G20_OPEN_CMD` 同值。
|
||
|
||
### 4.6 PLAN 名义路径
|
||
|
||
```text
|
||
_on_nominal(msg)
|
||
incoming = {name: position}
|
||
缺主动关节 → 与旧 goal 合并
|
||
goal = complete_g20_mimic_positions(merged)
|
||
track_nominal = {主动关节: goal[...]}
|
||
```
|
||
|
||
注意:Isaac 回调里**直接改 `self.goal`**,没有 MuJoCo 的 `_set_goal_only` + Lock;因为主循环同线程 `spin_once`,不存在并发写 Articulation。
|
||
|
||
---
|
||
|
||
## 5. 驱动模式:`set_dof_gains` 与 `apply_targets`
|
||
|
||
### 5.1 初始化时增益怎么设(`main` 里)
|
||
|
||
**默认位置硬写:**
|
||
|
||
```python
|
||
art.set_dof_gains(stiffnesses=zeros(n), dampings=zeros(n))
|
||
```
|
||
|
||
零增益,避免 PD 与 `set_dof_positions` 抢控制。
|
||
|
||
**`--pd-drive`:**
|
||
|
||
```python
|
||
# 主动关节
|
||
拇指四关节: stiff/damp/effort = 10 / 3 / 6
|
||
其余主动: stiff/damp/effort = 15 / 2.5 / 8
|
||
art.set_dof_gains(...); art.set_dof_max_efforts(...)
|
||
|
||
# mimic 关节强制 0 增益 —— 否则与 PhysX mimic 冲突 → Invalid PhysX transform
|
||
art.set_dof_gains(0, 0, dof_indices=mimic_idx)
|
||
```
|
||
|
||
`--hand-gravity` 仅在 PD 下生效:`set_link_enabled_gravities(True)`。
|
||
|
||
### 5.2 `apply_targets()` — 每物理步的核心
|
||
|
||
```python
|
||
def apply_targets(self):
|
||
self._slew_applied_toward_goal() # goal → applied
|
||
|
||
if use_physics_drive:
|
||
name_iter = (n for n in applied if n in actuated) # 只主动
|
||
else:
|
||
name_iter = applied.keys() # 含 mimic
|
||
|
||
positions = [...]
|
||
if use_physics_drive:
|
||
art.set_dof_position_targets(positions, dof_indices=indices)
|
||
else:
|
||
art.set_dof_positions(positions, dof_indices=indices)
|
||
art.set_dof_velocities(zeros(...), dof_indices=indices)
|
||
```
|
||
|
||
| 模式 | API | 写哪些关节 |
|
||
|------|-----|------------|
|
||
| 位置硬写(默认) | `set_dof_positions` + `set_dof_velocities(0)` | 全部(含软件算好的 mimic) |
|
||
| PD | `set_dof_position_targets` | 仅主动 16 个 |
|
||
|
||
### 5.3 `_slew_applied_toward_goal()`
|
||
|
||
```python
|
||
max_step = self.slew_rad # 默认 0 → 每步直达 goal
|
||
margin = LIMIT_MARGIN_RAD # 0.05,离限位留边
|
||
for name, g in goal.items():
|
||
g = clip(g, lo+margin, hi-margin)
|
||
if max_step <= 0:
|
||
applied[name] = g
|
||
else:
|
||
applied[name] = a ± min(|g-a|, max_step)
|
||
applied = complete_g20_mimic_positions(applied)
|
||
```
|
||
|
||
单位是 **每物理步弧度**,不是每秒。
|
||
物理 dt=1/120,默认每 render 16 步 → 墙钟约 16/120 ≈ 0.133 s 一帧渲染。
|
||
|
||
### 5.4 为何硬写后还要「再钉一次」?
|
||
|
||
主循环:
|
||
|
||
```python
|
||
bridge.apply_targets()
|
||
SimulationManager.step() # PhysX 可能轻微挪动 DOF
|
||
bridge.publish_state()
|
||
if not use_pd:
|
||
bridge.apply_targets() # 再钉回 applied
|
||
```
|
||
|
||
接触/积分会让硬写姿态漂移;步进后再写一遍,保证跟指令贴。
|
||
|
||
---
|
||
|
||
## 6. 异常恢复与状态发布
|
||
|
||
### 6.1 `recover_if_invalid()`
|
||
|
||
```python
|
||
pos = art.get_dof_positions()
|
||
if not all finite:
|
||
goal = applied = open_full
|
||
set_dof_positions(open)
|
||
return True
|
||
```
|
||
|
||
PD 爆炸或数值坏掉时复位张开姿态。
|
||
|
||
### 6.2 `publish_state()`
|
||
|
||
每物理步调用(不是独立 50Hz 定时器):
|
||
|
||
```python
|
||
pos = art.get_dof_positions()
|
||
vel = art.get_dof_velocities() # 失败则填 0
|
||
pub → /sim/isaac/g20/left/joint_state
|
||
```
|
||
|
||
`legacy_pub` 恒为 `None`,不会发 `/cb_left_hand_state`;
|
||
`urdf_positions_to_sdk_range` 已实现,预留反算 0~255。
|
||
|
||
### 6.3 `tracking_error()`
|
||
|
||
对 `track_nominal` 里每个主动关节:`|q_sim - q_nominal|`,返回 `(max, mean)`。
|
||
|
||
---
|
||
|
||
## 7. 自碰与防穿模(Isaac 怎么做)
|
||
|
||
**没有** MuJoCo 的 `collision_project`。手段是:
|
||
|
||
1. **USD `instances.usda`**:选择性开碰撞(掌心 + 指尖 ON,近端/中节多数 OFF)。
|
||
2. **`physx.usda`**:`enabledSelfCollisions = 1`。
|
||
3. **运行时**:`art.set_enabled_self_collisions([True])`(`--no-self-collision` 关)。
|
||
|
||
因此:
|
||
|
||
- 硬写时 PhysX **推不开** 被写死的 DOF → 仍可能视觉穿模。
|
||
- PD 模式下接触力可挡一部分,但跟手变差。
|
||
- 这与 MuJoCo「硬写前投影」是不同哲学。
|
||
|
||
---
|
||
|
||
## 8. 完整时序例子:收到半握命令后的一物理步
|
||
|
||
假设默认硬写、`slew_rad=0`。
|
||
|
||
```text
|
||
1. rclpy.spin_once
|
||
→ _on_cmd_u8 → sdk_range_to_full_urdf_positions
|
||
→ set_nominal_targets → goal = 半握全关节弧度
|
||
|
||
2. apply_targets()
|
||
→ _slew_applied_toward_goal: applied = goal(瞬达)
|
||
→ set_dof_positions(全部关节含 mimic)
|
||
→ set_dof_velocities(0)
|
||
|
||
3. SimulationManager.step()
|
||
→ PhysX 积分(硬写下姿态应贴近刚写的值)
|
||
|
||
4. recover_if_invalid() # 通常 False
|
||
|
||
5. publish_state()
|
||
→ 读 get_dof_positions → 发 /sim/isaac/...
|
||
|
||
6. apply_targets() 再钉一次(硬写模式)
|
||
```
|
||
|
||
若 `slew_rad=0.02`:每物理步每个关节最多走 0.02 rad,需多步才到 goal。
|
||
|
||
---
|
||
|
||
## 9. `--track-test` 路径
|
||
|
||
```text
|
||
max_steps 默认 480(若未指定)
|
||
每 160 物理步切换一次目标:
|
||
phase 偶: open_hand_command("G20")
|
||
phase 奇: [80.0]*20
|
||
set_nominal_targets(sdk_range_to_full_urdf_positions(...))
|
||
|
||
循环结束后再 settle 240 步:
|
||
apply_targets + SimulationManager.step
|
||
|
||
tracking_error() → max <= 0.02 → PASS,否则 exit 2
|
||
```
|
||
|
||
与 MuJoCo track-test(open→半握、关投影、slew=0)流程类似,但步进计数和切换方式不同。
|
||
|
||
---
|
||
|
||
## 10. 常用 Isaac / PhysX API 速查(本工程用到的)
|
||
|
||
| API | 用途 |
|
||
|-----|------|
|
||
| `SimulationApp` | 启动 Isaac 运行时 |
|
||
| `stage_utils.add_reference_to_stage` | 挂 USD |
|
||
| `prim_utils.set_prim_variants` | 选 Physics=physx |
|
||
| `Articulation(path)` | 包装手的 DOF |
|
||
| `set_dof_positions` / `set_dof_position_targets` | 硬写 / PD 目标 |
|
||
| `set_dof_gains` / `set_dof_max_efforts` | PD 参数 |
|
||
| `set_enabled_self_collisions` | 自碰开关 |
|
||
| `get_dof_positions` / `get_dof_velocities` / `get_dof_limits` | 读状态 |
|
||
| `SimulationManager.set_physics_dt` / `step` | 物理节拍 |
|
||
| `RenderingManager.render` | 渲染一帧 |
|
||
| `set_camera_view` | 对比视角 |
|
||
|
||
---
|
||
|
||
## 11. CLI 模式对照
|
||
|
||
| 参数 | 效果 |
|
||
|------|------|
|
||
| (默认) | 订 `/g20/cb_left_hand_control_cmd`,位置硬写,slew=0,自碰 ON |
|
||
| `--plan-nominal` | 订名义弧度话题,`use_cmd_u8=False` |
|
||
| `--legacy-sdk` | 订 `/cb_left_hand_control_cmd` |
|
||
| `--pd-drive` | `set_dof_position_targets` + 主动关节增益 |
|
||
| `--hand-gravity` | 仅 PD 下开手部重力 |
|
||
| `--no-self-collision` | 关自碰(会穿模) |
|
||
| `--slew-rad N` | 每物理步最多转 N rad |
|
||
| `--phys-per-render N` | 每渲染帧物理步数(默认 16) |
|
||
| `--track-test` | 内置跟踪验收 |
|
||
|
||
---
|
||
|
||
## 12. 改代码时最常动的位置
|
||
|
||
| 需求 | 改哪里 |
|
||
|------|--------|
|
||
| SDK→弧度、mimic、OPEN_CMD | `joint_mapping.py` |
|
||
| 默认硬写/PD 增益、拇指特殊增益 | `isaac_g20_plan_bridge.py` 顶部常量 + `main` 里 `set_dof_gains` |
|
||
| slew / 物理步 / 安装高度 | `DEFAULT_SLEW_RAD` / `DEFAULT_PHYS_PER_RENDER` / `HAND_MOUNT_*` |
|
||
| 相机 | `COMPARE_*` + `_compare_camera_eye_target` |
|
||
| 碰撞谁开谁关 | USD `payloads/instances.usda` |
|
||
| 自碰/求解器 | USD `payloads/Physics/physx.usda` |
|
||
| 话题名 | `TOPIC_*` 或 CLI |
|
||
|
||
---
|
||
|
||
## 13. 与 MuJoCo 技术详解的对照表
|
||
|
||
| 概念 | Isaac 函数/位置 | MuJoCo 函数/位置 |
|
||
|------|-----------------|------------------|
|
||
| 映射入口 | `joint_mapping.sdk_range_to_full_urdf_positions` | `g20_joints.sdk_range_to_full_urdf_positions` |
|
||
| 收命令 | `_on_cmd_u8` / `_on_nominal` | `_on_legacy_cmd` / `_on_nominal` |
|
||
| 更新目标 | `set_nominal_targets` / 直接改 `goal` | `_set_goal_only`(持锁) |
|
||
| 限速 | `_slew_applied_toward_goal`(rad/步) | `_slew_applied_unlocked`(rad/s) |
|
||
| 写入仿真 | `apply_targets` → `set_dof_*` | `_hard_write_pose_unlocked` → `mj_forward` |
|
||
| 防穿模 | USD collider + 自碰(无投影) | `_project_applied_no_penetration_unlocked` |
|
||
| 物理步进 | `SimulationManager.step` | 硬写不 step;PD 用 `mj_step` |
|
||
| 发状态 | 每物理步 `publish_state` | 50Hz timer `_publish_state` |
|
||
| 状态话题 | `/sim/isaac/g20/left/joint_state` | `/sim/mujoco/g20/left/joint_state` |
|
||
|
||
---
|
||
|
||
## 14. 读代码建议顺序
|
||
|
||
1. `run_g20_plan.sh` → 确认用的是哪份 `python.sh`
|
||
2. 文件顶层:`parse_args` → `SimulationApp` → import
|
||
3. `main()` 前半:`_load_g20` → `_fix_base` → gains
|
||
4. `G20PlanBridge.__init__` + `_on_cmd_u8`
|
||
5. `apply_targets` + 主循环里「写→step→再钉」
|
||
6. 需要改防穿模时去看 USD `instances.usda`,而不是在桥里找 project 函数
|
||
|
||
资产与改动清单仍见 [`G20_Isaac_实现说明.md`](./G20_Isaac_实现说明.md)。
|