#!/usr/bin/env python3 """Prepare one locked standing actor candidate; publication is gated separately.""" import argparse import json import shutil from pathlib import Path from runtime_common import write_json SELECTED_SHA256 = "f2197339dea3622ed92b82fdddff3400f2a82f3ca4d4fd3be6c72b45f1a77704" def prepare(run, output): import torch from SkillFactory.legs.frozen_standing import SHAPES, Actor, FrozenStanding, file_hash from SkillFactory.legs.standing_ppo import StandingActorCritic, features from SkillFactory.legs.urdf import forward, multiply, pose torch.set_num_threads(4) selected = run / "checkpoints/best_validation.pt" if file_hash(selected) != SELECTED_SHA256: raise ValueError("Fixed selected checkpoint changed; do not reselect from holdout") result = json.loads((run / "result.json").read_text()) selection = json.loads((run / "selection.json").read_text()) if not result["formal_training_complete"] or result["update"] != 2048 or selection["update"] != 1024: raise ValueError("Incomplete training/wrong validation selection") checkpoint = torch.load(selected, map_location="cpu", weights_only=True) if checkpoint["update"] != 1024 or checkpoint["identity"] != result["identity"]: raise ValueError("Checkpoint identity mismatch") for name, digest in checkpoint["identity"]["sources"].items(): if file_hash(name) != digest: raise ValueError(f"Training source modified: {name}") task_path = Path("configs/standing/go2_fl_support_v0.json") task = json.loads(task_path.read_text()) support = Path(task["support_asset"]) / "support.json" bundle = json.loads((Path(task["source_bundle"]) / "manifest.json").read_text()) nominal = [j["q_default"] for j in bundle["spec"]["joints"]] poses, _ = forward(bundle["tree"], dict(zip(bundle["joint_order"], nominal))) foot = bundle["foot"] height = foot["radius"] - multiply(poses[foot["body"]], pose(foot["xyz"], foot["rpy"]))[2][3] output.mkdir(parents=True, exist_ok=False) binding = { "task_sha256": file_hash(task_path), "support_sha256": file_hash(support), "source_manifest_sha256": file_hash(Path(task["source_bundle"]) / "manifest.json"), "training_identity": checkpoint["identity"], "checkpoint_sha256": SELECTED_SHA256, "height_m": height, "nominal": nominal, "action_scale_rad": task["action_scale_rad"], "soft_limits": [j["soft_limits"] for j in bundle["joints"]], "joint_order": bundle["joint_order"], "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": json.loads(support.read_text()), } write_json(output / "binding.json", binding) for source, target in ( (task_path, "task.json"), (run / "config.json", "training.json"), (run / "selection.json", "selection.json"), (Path("source/SkillFactory/SkillFactory/legs/frozen_standing.py"), "frozen_standing.py"), (Path(task["source_bundle"]) / "LICENSE", "ASSET_LICENSE"), ): shutil.copyfile(source, output / target) actor = Actor(checkpoint["model_config"]["hidden"]) actor.load_state_dict( {k: v for k, v in checkpoint["model"].items() if k.startswith("actor.") or k in ("offset", "scale")} ) torch.save(actor.state_dict(), output / "actor.pt") obs = {k: torch.zeros(1, *shape) for k, shape in SHAPES.items()} obs["joint_state"][0, :, 0] = torch.tensor(nominal) obs["base_state"][0, 2], obs["base_state"][0, 6], obs["goal"][0, 0] = height, 1, height torch.save(obs, output / "example.pt") (output / "requirements.txt").write_text( "# Python3.12; CPU or CUDA PyTorch2.10.0 as separately validated\ntorch==2.10.0\n" ) (output / "README.md").write_text( "# Go2 FL guided loaded stand 1.0.0\n\n" "Inference requires Python3.12/Torch2.10.0 only, no Kit/project/training checkpoint. " "Load FrozenStanding(directory, manifest_sha256=, " "task_sha256=, support_sha256=, 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.\n\n" "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.\n\n" "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.\n\n" "## Exact observation ABI (all float32, same device, nonempty batch B)\n\n" "- joint_state[B,3,3]: declared joint order; q(rad), dq(rad/s), previous normalized action.\n" "- 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.\n" "- 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.\n" "- 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).\n" "- guide_wrench[B,6]: actual passive guide force/torque in WORLD frame, mount origin, N/Nm; Fz exactly0.\n" "- goal[B,1]: fixed nominal payload mounting-origin height in m, binding.height_m.\n" "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.\n" ) manifest = { "format": "skillfactory.standing-bundle.v1", "contract": "go2-fl-loaded-stand-v0", "version": "1.0.0", "stage": "candidate", "hidden": checkpoint["model_config"]["hidden"], "checkpoint_sha256": SELECTED_SHA256, "files": {p.name: file_hash(p) for p in sorted(output.iterdir()) if p.is_file()}, } write_json(output / "manifest.json", manifest) frozen = FrozenStanding( output, manifest_sha256=file_hash(output / "manifest.json"), task_sha256=binding["task_sha256"], support_sha256=binding["support_sha256"], ) model = StandingActorCritic(**checkpoint["model_config"]).eval() model.load_state_dict(checkpoint["model"]) gen = torch.Generator().manual_seed(91000) worst = 0.0 for batch in (1, 16, 64, 256): sample = {k: v.expand(batch, *v.shape[1:]).clone() for k, v in obs.items()} for key in ("joint_state", "mount_wrench", "foot_state"): sample[key] += 0.05 * torch.randn(sample[key].shape, generator=gen) with torch.inference_mode(): difference = (model.act(features(sample), deterministic=True)[0] - frozen.act(sample)).abs().max().item() worst = max(worst, difference) if worst > 1e-6: raise ValueError("Actor extraction mismatch") write_json( output.parent / "candidate_lock.json", { "passed": True, "accepted_skill": False, "checkpoint_sha256": SELECTED_SHA256, "selected_update": 1024, "manifest_sha256": file_hash(output / "manifest.json"), "core_files": manifest["files"], "actor_extraction_max_error": worst, }, ) def main(): parser = argparse.ArgumentParser(description=__doc__) parser.add_argument("--run", type=Path, default=Path("logs/loaded_stand/train_v1")) parser.add_argument("--output", type=Path, required=True) args = parser.parse_args() prepare(args.run, args.output) if __name__ == "__main__": main()