"""Client-side test of the pi0.5 MMB policy server (shared interface): sends real dataset frames (raw 480x640 RGB head / wrist images, the stored canonical state, robot_id, task string) and checks the response format; compares the predicted chunk with the training target (expert targets + recorded commands) and decodes it with the real base pose. Usage: OPENPI_VIDEO_BACKEND=pyav HF_LEROBOT_HOME=/data/changyi/mmb/lerobot python test_server_client.py [--port 8000] [--n 20]""" import argparse import json import random import time from pathlib import Path import numpy as np import pyarrow.parquet as pq from lerobot.common.datasets.lerobot_dataset import LeRobotDataset from openpi_client import websocket_client_policy from openpi.policies import mmb_canonical_action as ca ROOT = Path("/data/changyi/mmb/merged_v2t") NAMES = {0: "go2_arm", 1: "sawyer_umi", 2: "stretch3"} def to_uint8_hwc(img): x = np.asarray(img) if x.shape[0] == 3: x = np.transpose(x, (1, 2, 0)) if x.dtype != np.uint8: x = np.clip(np.round(x * 255), 0, 255).astype(np.uint8) return x def main(): ap = argparse.ArgumentParser(); ap.add_argument("--port", type=int, default=8000); ap.add_argument("--n", type=int, default=20) a = ap.parse_args() client = websocket_client_policy.WebsocketClientPolicy("localhost", a.port) meta = client.get_server_metadata() print("server metadata:", json.dumps(meta, default=str), flush=True) H = int(meta["action_horizon"]) ds = LeRobotDataset("mmb/v2t", video_backend="pyav") eps = [json.loads(l) for l in open(ROOT / "meta/episodes.jsonl")] src = [json.loads(l) for l in open(ROOT / "meta/source_map.jsonl")] offsets = np.concatenate([[0], np.cumsum([e["length"] for e in eps])]) random.seed(7) by = {} for i, s in enumerate(src): by.setdefault(s["robot"], []).append(i) res = {r: {"pos_cm": [], "rot_deg": [], "grip_ok": [], "base_err": [], "ms": [], "dec_pos_cm": []} for r in ca.ROBOTS} fmt_ok = True for robot in ca.ROBOTS: for ep in random.sample(by[robot], a.n): L = eps[ep]["length"]; t = random.randrange(L) item = ds[int(offsets[ep] + t)] tab = pq.read_table(ROOT / f"data/chunk-{ep // 1000:03d}/episode_{ep:06d}.parquet", columns=["observation.state_canonical", "action.command", "action.eef_target", "observation.base_pose", "actions"]) S, Cm, G, B, A = (np.asarray(tab.column(c).to_pylist(), np.float64) for c in tab.column_names) rid = {"go2_arm": 0, "sawyer_umi": 1, "stretch3": 2}[robot] obs = {"observation/image": to_uint8_hwc(item["observation.images.head"]), "observation/wrist_image": to_uint8_hwc(item["observation.images.wrist"]), "observation/state": S[t].astype(np.float32), "robot_id": rid, "prompt": item["task"]} assert obs["observation/image"].shape == (480, 640, 3), obs["observation/image"].shape t0 = time.time(); out = client.infer(obs); ms = (time.time() - t0) * 1000 act = np.asarray(out["actions"]) fmt_ok &= act.shape == (H, 14) and act.dtype == np.float32 and bool(np.all(act[:, ~ca.ACTION_MASK[robot]] == 0)) idx = np.minimum(t + np.arange(H), L - 1) gt = ca.action_chunk_from_targets(robot, S[t], G[idx], Cm[idx]) r = res[robot] r["ms"].append(ms) r["pos_cm"].append(np.linalg.norm(act[:, ca.A_POS] - gt[:, ca.A_POS], axis=-1) * 100) Rp, Rg = ca.rot6d_to_rotmat(act[:, ca.A_ROT]), ca.rot6d_to_rotmat(gt[:, ca.A_ROT]) r["rot_deg"].append(ca.rot_angle_deg(np.einsum("hji,hjk->hik", Rg, Rp))) r["grip_ok"].append(((act[:, ca.A_GRIP] > 0.5) == (gt[:, ca.A_GRIP] > 0.5)).astype(float)) m = ca.ACTION_MASK[robot][:4] r["base_err"].append(np.abs(act[:, :4][:, m] - gt[:, :4][:, m]).max(1)) dec = ca.decode_action_chunk(robot, act, S[t], B[idx]) # world EEF target vs recorded expert target o = 4 if robot == "go2_arm" else 3 r["dec_pos_cm"].append(np.linalg.norm(dec[:, o:o + 3] - A[idx][:, o:o + 3], axis=-1) * 100) summary = {"format_ok": bool(fmt_ok), "horizon": H} for robot, r in res.items(): c = {k: np.concatenate([np.ravel(x) for x in v]) for k, v in r.items()} summary[robot] = {"requests": len(r["ms"]), "latency_ms_p50_max": [round(float(np.median(c["ms"])), 1), round(float(c["ms"].max()), 1)], "eef_pos_err_cm_p50_p90": [round(float(np.percentile(c["pos_cm"], q)), 2) for q in (50, 90)], "eef_rot_err_deg_p50_p90": [round(float(np.percentile(c["rot_deg"], q)), 2) for q in (50, 90)], "gripper_agreement": round(float(c["grip_ok"].mean()), 3), "base_cmd_abs_err_p50_p90": [round(float(np.percentile(c["base_err"], q)), 3) for q in (50, 90)], "decoded_world_target_vs_recorded_cm_p50": round(float(np.median(c["dec_pos_cm"])), 2), # same summary as the GR00T test: worst step of each chunk, median over requests "chunk_max_median_pos_cm_rot_deg_base": [round(float(np.median([x.max() for x in r[k]])), 3) for k in ("pos_cm", "rot_deg", "base_err")], "chunks_with_all_gripper_steps_correct": round(float(np.mean([x.min() for x in r["grip_ok"]])), 3)} print(json.dumps(summary, indent=1), flush=True) if __name__ == "__main__": main()