"""Serve a trained dexhand pi0.5 checkpoint over openpi's websocket protocol (openpi venv; no data needed). python train/serve.py --checkpoint / [--port 8000] [--max-token-len 288] [--horizon 50] [--right-wrist black] The checkpoint dir must hold params/ and assets/dual_marvin_wuji/norm_stats.json (as saved by training). Request (one dict, unbatched): observation/image (head, HxWx3 uint8), observation/left_wrist_image, observation/right_wrist_image, observation/state (54,), prompt (sub-task instruction). Reply: actions (T, 54) absolute joint targets in the data order (arm 7 + hand 20 finger-major, per side), via the training config's output transforms (de-interleave, unnormalise, arm deltas -> absolute). """ import argparse import logging import pathlib import sys sys.path.insert(0, str(pathlib.Path(__file__).resolve().parent)) class BlackRightWrist: """Policy wrapper: the right wrist image is replaced by zeros before inference.""" def __init__(self, policy): self._policy = policy def infer(self, obs): import numpy as np obs = dict(obs) obs["observation/right_wrist_image"] = np.zeros_like(np.asarray(obs["observation/right_wrist_image"])) return self._policy.infer(obs) def __getattr__(self, name): return getattr(self._policy, name) def main(): p = argparse.ArgumentParser() p.add_argument("--checkpoint", required=True) p.add_argument("--port", type=int, default=8000) p.add_argument("--horizon", type=int, default=50) p.add_argument("--max-token-len", type=int, default=288) # as trained (run.py; mid-train v1 log) p.add_argument("--num-steps", type=int, default=10, help="flow-matching denoising steps") p.add_argument("--right-wrist", default="keep", choices=["keep", "black"], help="black: replace the right wrist image with zeros (models trained before 2026-10-26 saw a " "right wrist camera on the BACK of the hand, filming the ceiling; the real one looks at the table)") a = p.parse_args() logging.basicConfig(level=logging.INFO, force=True) import dexhand_openpi as dx from openpi.policies import policy_config from openpi.serving import websocket_policy_server cfg = dx.get_config("unused", T=a.horizon, max_token_len=a.max_token_len) policy = policy_config.create_trained_policy(cfg, a.checkpoint, sample_kwargs={"num_steps": a.num_steps}) if a.right_wrist == "black": policy = BlackRightWrist(policy) meta = {"checkpoint": str(a.checkpoint), "horizon": a.horizon, "action_dim": 54, "right_wrist": a.right_wrist} logging.info("serving %s on port %d", a.checkpoint, a.port) websocket_policy_server.WebsocketPolicyServer(policy=policy, host="0.0.0.0", port=a.port, metadata=meta).serve_forever() if __name__ == "__main__": main()