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
+201
View File
@@ -0,0 +1,201 @@
#!/usr/bin/env python3
"""Select validation checkpoint, export candidate, or finalize accepted immutable release."""
import argparse
import json
import shutil
import tempfile
from pathlib import Path
from runtime_common import critical_log_lines, write_json
def manifest(directory, cfg, status):
from SkillFactory.legs.offline import sha256
from stage2_acceptance import package_version
files = {
str(path.relative_to(directory)): sha256(path)
for path in sorted(directory.rglob("*"))
if path.is_file() and path.name != "manifest.json" and "__pycache__" not in path.parts
}
write_json(
directory / "manifest.json",
{
"format": "skillfactory.skillbundle.v1",
"interface": "go2-fl-reach-leg-token-v1",
"version": package_version(directory, cfg),
"status": status,
"frozen": True,
"hidden": cfg["training"]["hidden"],
"dependencies": {"python": "3.12.x", "torch": "2.10.0 (CPU validated; CUDA evaluated separately)"},
"scope": cfg["scope"],
"files": files,
},
)
USAGE = """# Go2 FL Reach SkillBundle @VERSION@
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-@VERSION@')
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。
"""
def export(args, cfg):
import torch
from SkillFactory.legs.offline import canonical_hash, sha256
train_result = json.loads((args.run / "train/result.json").read_text())
if not train_result["training_complete"] or train_result["updates"] != cfg["training"]["updates"]:
raise ValueError("Formal training not completed")
candidates = []
for step in cfg["selection"]["candidate_updates"]:
path = args.run / f"validation_{step:06d}/result.json"
row = json.loads(path.read_text())
weights = args.run / f"train/step_{step:06d}.pt"
if row["weight_sha256"] != sha256(weights) or row["config_sha256"] != canonical_hash(cfg):
raise ValueError("Validation provenance mismatch")
if row["passed"] and all(row["gates"].values()):
log_path = args.run / f"validation_{step:06d}.log"
if not log_path.is_file() or critical_log_lines(log_path):
raise ValueError("Missing/critical validation log")
candidates.append((row["policy"]["mean_error_m"], step, weights))
if not candidates:
raise ValueError("No candidate passed pre-registered validation gates; no release")
_, step, weights = min(candidates)
args.output.mkdir(parents=True, exist_ok=False)
if getattr(args, "release_version", None):
write_json(
args.output / "release.json",
{
"format": "skillfactory.release.v1",
"version": args.release_version,
"training_config_sha256": canonical_hash(cfg),
},
)
from stage2_acceptance import package_version
version = package_version(args.output, cfg)
checkpoint = torch.load(weights, weights_only=True, map_location="cpu")
torch.save(checkpoint["learner"]["models"]["actor"], args.output / "actor.pt")
for name in ("binding.json", "example_tokens.pt"):
shutil.copy2(args.run / "data" / name, args.output / name)
shutil.copy2(
Path(__file__).resolve().parents[1] / "source/SkillFactory/SkillFactory/legs/frozen_policy.py",
args.output / "frozen_policy.py",
)
shutil.copy2(Path(cfg["bundle"]) / "LICENSE", args.output / "ASSET_LICENSE")
shutil.copy2(args.config, args.output / "config.json")
(args.output / "README.md").write_text(USAGE.replace("@VERSION@", version))
(args.output / "requirements.txt").write_text("torch==2.10.0\n")
selection = {
"selected_update": step,
"checkpoint_sha256": sha256(weights),
"actor_sha256": sha256(args.output / "actor.pt"),
"rule": cfg["selection"]["rule"],
"selection_seeds": [cfg["selection"]["seed"] + i for i in range(cfg["selection"]["batches"])],
"config_sha256": canonical_hash(cfg),
"training_identity": checkpoint["identity"],
}
write_json(args.output / "selection.json", selection)
manifest(args.output, cfg, "candidate")
print(json.dumps(selection, indent=2))
def finalize(args, cfg):
from stage2_acceptance import evaluate_acceptance, evaluate_release, package_version, require_acceptance
from verify_frozen_leg import verify_package
if args.output.exists():
raise ValueError("Immutable version already exists; publish a new version")
if getattr(args, "release_version", None) not in (None, package_version(args.candidate, cfg)):
raise ValueError("Finalization cannot change the tested candidate version")
# Mandatory engineering/resume/data/training/selection/physics/load gates are shared
# with the delivery checker. Do not create any public package before they pass.
require_acceptance(evaluate_acceptance(args.run, args.candidate, cfg))
if not getattr(args, "python", None):
raise ValueError("Finalization requires --python pointing to an isolated inference interpreter")
args.output.parent.mkdir(parents=True, exist_ok=True)
lock = args.output.parent / f".{args.output.name}.publish.lock"
with lock.open("x"):
try:
with tempfile.TemporaryDirectory(prefix=f".{args.output.name}-", dir=args.output.parent) as tmp:
staged = Path(tmp) / "package"
shutil.copytree(args.candidate, staged, ignore=shutil.ignore_patterns("__pycache__"))
evidence = staged / "evidence"
evidence.mkdir(exist_ok=True)
for mode in ("test", "boundary", "stability"):
shutil.copy2(args.run / mode / "result.json", evidence / f"{mode}.json")
for name in ("standalone.json", "resume_exact.json", "engineering_checks.json"):
shutil.copy2(args.run / name, evidence / name)
shutil.copy2(args.candidate / "manifest.json", evidence / "tested_candidate_manifest.json")
for step in cfg["selection"]["candidate_updates"]:
shutil.copy2(
args.run / f"validation_{step:06d}/result.json", evidence / f"validation_{step:06d}.json"
)
for name in ("train/result.json", "data/dataset.json"):
shutil.copy2(args.run / name, evidence / name.replace("/", "_"))
for name in ("provenance.json", "inference_dependencies.txt"):
if (args.run / name).is_file():
shutil.copy2(args.run / name, evidence / name)
# Final metadata changes the manifest hash: independently load these exact
# staged bytes too, before the accepted version becomes publicly visible.
manifest(staged, cfg, "accepted")
standalone = verify_package(staged, args.python, args.run / "release_standalone.json")
require_acceptance(evaluate_release(args.run, staged, cfg, standalone))
if args.output.exists():
raise ValueError("Immutable version already exists; publish a new version")
staged.rename(args.output)
finally:
lock.unlink(missing_ok=True)
print(f"Accepted immutable package: {args.output}")
def main():
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--config", type=Path, required=True)
parser.add_argument("--run", type=Path, required=True)
parser.add_argument("--output", type=Path, required=True)
parser.add_argument("--candidate", type=Path, help="If supplied, finalize this candidate after independent gates")
parser.add_argument("--python", type=Path, help="Required for finalization: isolated CPU inference interpreter")
parser.add_argument(
"--release-version", help="Candidate release version; does not modify the locked training config"
)
args = parser.parse_args()
cfg = json.loads(args.config.read_text())
finalize(args, cfg) if args.candidate else export(args, cfg)
if __name__ == "__main__":
main()