import os import numpy as np import requests import argparse from PIL import Image from pathlib import Path from typing import Any, Dict, Optional from dataclasses import dataclass from numpy.lib.format import descr_to_dtype, dtype_to_descr from base64 import b64decode, b64encode # ============ Serialization Utils ============ def numpy_serialize(o): if isinstance(o, (np.ndarray, np.generic)): data = o.data if o.flags["C_CONTIGUOUS"] else o.tobytes() return { "__numpy__": b64encode(data).decode(), "dtype": dtype_to_descr(o.dtype), "shape": o.shape, } msg = f"Object of type {o.__class__.__name__} is not JSON serializable" raise TypeError(msg) def numpy_deserialize(dct): if "__numpy__" in dct: np_obj = np.frombuffer(b64decode(dct["__numpy__"]), descr_to_dtype(dct["dtype"])) return np_obj.reshape(shape) if (shape := dct["shape"]) else np_obj[0] return dct def convert_numpy_in_dict(data, func): if isinstance(data, dict): if "__numpy__" in data: return func(data) return {key: convert_numpy_in_dict(value, func) for key, value in data.items()} elif isinstance(data, list): return [convert_numpy_in_dict(item, func) for item in data] elif isinstance(data, (np.ndarray, np.generic)): return func(data) else: return data # ============ Message Classes ============ class RequestMessage: def __init__(self, image: Dict[str, Any], instruction: str, history: Dict[str, Any], state: Dict[str, Any], gt_action, dataset_name: str): self.image = image self.instruction = instruction self.history = history self.state = state self.gt_action = gt_action self.dataset_name = dataset_name def serialize(self): from datetime import datetime msg = { "image": self.image, "instruction": self.instruction, "history": self.history, "state": self.state, "condition": {}, # Not used by ACT but required by server "gt_action": self.gt_action, "dataset_name": self.dataset_name, "timestamp": str(datetime.now()).replace(" ", "_").replace(":", "-") } return convert_numpy_in_dict(msg, numpy_serialize) class ResponseMessage: def __init__(self, action: np.ndarray, err: float): self.action = action self.err = err @classmethod def deserialize(cls, response: Dict[str, Any]): response = convert_numpy_in_dict(response, numpy_deserialize) return cls(action=response["action"], err=response["err"]) # ============ HTTP Client ============ class ACTActionClient: def __init__(self, server_ip: str = "127.0.0.1", server_port: int = 22089): self.server_ip = server_ip self.server_port = server_port def query_action(self, image_dict: Dict[str, np.ndarray], instruction: str, state_dict: Dict[str, np.ndarray], gt_action: Optional[np.ndarray] = None, history: Optional[Dict[str, Any]] = None, dataset: str = "grasp") -> tuple[np.ndarray, float]: if history is None: history = {k: [] for k in image_dict.keys()} if gt_action is None: gt_action = [] request = RequestMessage(image_dict, instruction, history, state_dict, gt_action, dataset) response = requests.post( f"http://{self.server_ip}:{self.server_port}/act", json=request.serialize() ) response = ResponseMessage.deserialize(response.json()) return response.action, response.err def health_check(self) -> bool: try: response = requests.get(f"http://{self.server_ip}:{self.server_port}/health") return response.json().get("status") == "ok" except Exception as e: print(f"Health check failed: {e}") return False # ============ Main Evaluation ============ def main(): parser = argparse.ArgumentParser(description="Act Client for testing Action Chunking Transformer server") parser.add_argument("--ip", type=str, default="127.0.0.1", help="Server IP address") parser.add_argument("--port", type=int, default=22085, help="Server port") parser.add_argument("--dataset-path", type=str, default="/hfm/data/real_teleop_g1/lerobot/Pick_bottle_and_turn_and_pour_into_cup", help="Path to LeRobot dataset") parser.add_argument("--episode-idx", type=int, default=0, help="Episode index to evaluate") args = parser.parse_args() np.set_printoptions(precision=4, suppress=True) # 1. Create client client = ACTActionClient(args.ip, args.port) # Health check print(f"Connecting to server at {args.ip}:{args.port}...") if not client.health_check(): print("❌ Server is not responding. Please start act_serve.py first.") return print("✅ Server is healthy!") # 2. Load dataset print(f"\nLoading dataset from {args.dataset_path}...") from lerobot.common.datasets.lerobot_dataset import LeRobotDataset dataset_name = Path(args.dataset_path).name raw_dataset = LeRobotDataset(dataset_name, args.dataset_path) # Helper function - LeRobot images are in [0, 1] range (not normalized to [-1, 1]) def tensor_to_numpy(x): """Convert tensor [0, 1] to uint8 image""" return (x.float().clamp(0, 1) * 255.0).permute(1, 2, 0).cpu().numpy().astype(np.uint8) # 3. Evaluate one episode eps_idx = args.episode_idx from_idx = raw_dataset.episode_data_index["from"][eps_idx].item() to_idx = raw_dataset.episode_data_index["to"][eps_idx].item() print(f"\n{'='*50}") print(f"Evaluating Episode {eps_idx}") print(f"Frame range: {from_idx} - {to_idx}") print(f"Episode length: {to_idx - from_idx} frames") print(f"{'='*50}\n") all_action_l1_errs = [] l2_xyz_errors = [] l2_rpy_errors = [] l2_gripper_errors = [] for idx in range(from_idx, to_idx - 1): frame = raw_dataset[idx] # Prepare image - LeRobot images are already in [0, 1] range image = tensor_to_numpy(frame["observation.images.front_stereo_left"]) # Prepare state obs = frame["observation.state"].numpy() state_dict = { "proprio_joint_positions": obs[:8], "proprio_eef_pose": obs[-7:] } # Prepare instruction and GT action instruction = frame["task"] gt_action = frame["action"].numpy() # Send request observations = {"front_stereo_left": image} try: pred_actions, err = client.query_action( observations, instruction, state_dict, gt_action=gt_action ) except Exception as e: print(f"❌ Request failed at frame {idx}: {e}") continue # Calculate errors (pred_actions is (Tp, Da), gt_action is (Da,)) # Only compare the first predicted action with gt_action action_l1_errs = np.abs(pred_actions[0] - gt_action) # Compare first action only avg_action_errors = action_l1_errs # (Da,) all_action_l1_errs.append(action_l1_errs[np.newaxis, :]) # Split into xyz, rpy, gripper avg_l1_action_err = np.split(avg_action_errors, [3, 6], axis=-1) xyz_err = np.linalg.norm(avg_l1_action_err[0]) rpy_err = np.linalg.norm(avg_l1_action_err[1]) gripper_err = np.linalg.norm(avg_l1_action_err[2]) metric = { "denorm_err_l1_xyz": xyz_err, "denorm_err_l1_rpy": rpy_err, "denorm_err_l1_gripper": gripper_err, } print(f"Frame {idx - from_idx:3d} | {metric}") l2_xyz_errors.append(xyz_err) l2_rpy_errors.append(rpy_err) l2_gripper_errors.append(gripper_err) # ============ Summary Statistics ============ print(f"\n{'='*50}") print("Episode Summary") print(f"{'='*50}") l2_xyz_errors = np.array(l2_xyz_errors) l2_rpy_errors = np.array(l2_rpy_errors) l2_gripper_errors = np.array(l2_gripper_errors) from we.utils import rmse print(f"XYZ Error: Mean={np.mean(l2_xyz_errors):.4f} ± {np.std(l2_xyz_errors):.4f}, RMSE={rmse(l2_xyz_errors):.4f}") print(f"RPY Error: Mean={np.mean(l2_rpy_errors):.4f} ± {np.std(l2_rpy_errors):.4f}, RMSE={rmse(l2_rpy_errors):.4f}") print(f"Gripper Error: Mean={np.mean(l2_gripper_errors):.4f} ± {np.std(l2_gripper_errors):.4f}, RMSE={rmse(l2_gripper_errors):.4f}") print(f"\nTotal frames evaluated: {len(l2_xyz_errors)}") # ============ Quality Assessment ============ print(f"\n{'='*50}") print("Quality Assessment") print(f"{'='*50}") xyz_mean = np.mean(l2_xyz_errors) rpy_mean = np.mean(l2_rpy_errors) if xyz_mean < 0.01: xyz_quality = "✅ Good (< 1cm)" elif xyz_mean < 0.03: xyz_quality = "⚠️ Acceptable (1-3cm)" else: xyz_quality = "❌ Poor (> 3cm)" if rpy_mean < 0.1: rpy_quality = "✅ Good (< 5.7°)" elif rpy_mean < 0.3: rpy_quality = "⚠️ Acceptable (5.7-17°)" else: rpy_quality = "❌ Poor (> 17°)" print(f"Position (XYZ): {xyz_quality} - Mean error: {xyz_mean*100:.2f} cm") print(f"Rotation (RPY): {rpy_quality} - Mean error: {np.degrees(rpy_mean):.2f}°") if __name__ == "__main__": main()