|
| 1 | +#!/usr/bin/env python3 |
| 2 | +"""Run Panthera single-arm real-robot inference through ModelClient. |
| 3 | +
|
| 4 | +The TCP model server side should expose an RPC method such as ``infer(obs)`` |
| 5 | +that returns either a single action [14] or an action chunk [T, 14]. For this |
| 6 | +single-arm robot, the first 7 dims are used as left_arm joint(6)+gripper(1), |
| 7 | +and the remaining right-arm dims are ignored. |
| 8 | +""" |
| 9 | + |
| 10 | +from __future__ import annotations |
| 11 | + |
| 12 | +import argparse |
| 13 | +from pathlib import Path |
| 14 | +import sys |
| 15 | +import time |
| 16 | +from typing import Any |
| 17 | + |
| 18 | +import numpy as np |
| 19 | + |
| 20 | +THIS_FILE = Path(__file__).resolve() |
| 21 | +DEPLOY_DIR = THIS_FILE.parent |
| 22 | +REPO_ROOT = DEPLOY_DIR.parent.parent |
| 23 | + |
| 24 | +for path in (DEPLOY_DIR, REPO_ROOT, REPO_ROOT / "src"): |
| 25 | + path_str = str(path) |
| 26 | + if path_str not in sys.path: |
| 27 | + sys.path.insert(0, path_str) |
| 28 | + |
| 29 | +from robot_client_server.model_client import ModelClient |
| 30 | +from my_robot.panthera_single import PantheraSingle |
| 31 | + |
| 32 | + |
| 33 | +ACT_STATE_DIM = 14 |
| 34 | +LEFT_ACTION_DIM = 7 |
| 35 | + |
| 36 | + |
| 37 | +def _color_to_chw(color: Any) -> np.ndarray: |
| 38 | + image = np.asarray(color) |
| 39 | + if image.ndim != 3 or image.shape[2] != 3: |
| 40 | + raise ValueError(f"expected HWC RGB/BGR image, got shape={image.shape}") |
| 41 | + return np.transpose(image.astype(np.uint8, copy=False), (2, 0, 1)) |
| 42 | + |
| 43 | + |
| 44 | +def _build_state(controller_data: dict[str, Any]) -> np.ndarray: |
| 45 | + left = controller_data["left_arm"] |
| 46 | + joint = np.asarray(left["joint"], dtype=np.float32).reshape(-1) |
| 47 | + if joint.shape != (6,): |
| 48 | + raise ValueError(f"left_arm.joint must be shape (6,), got {joint.shape}") |
| 49 | + gripper = np.asarray(left["gripper"], dtype=np.float32).reshape(1) |
| 50 | + state = np.zeros((ACT_STATE_DIM,), dtype=np.float32) |
| 51 | + state[:6] = joint |
| 52 | + state[6] = gripper[0] |
| 53 | + return state |
| 54 | + |
| 55 | + |
| 56 | +def build_observation(raw_data: Any, prompt: str) -> dict[str, Any]: |
| 57 | + controller_data, sensor_data = raw_data |
| 58 | + state = _build_state(controller_data) |
| 59 | + cam_high = _color_to_chw(sensor_data["cam_head"]["color"]) |
| 60 | + cam_wrist = _color_to_chw(sensor_data["cam_wrist"]["color"]) |
| 61 | + |
| 62 | + # Keep cam_wrist for single-arm checkpoints. Also provide the dual-arm keys as |
| 63 | + # aliases so older OpenPI/ALOHA wrappers that still look for wrist-left/right |
| 64 | + # can consume the same observation without changing this robot client. |
| 65 | + return { |
| 66 | + "state": state, |
| 67 | + "images": { |
| 68 | + "cam_high": cam_high, |
| 69 | + "cam_wrist": cam_wrist, |
| 70 | + "cam_right_wrist": cam_wrist, |
| 71 | + "cam_left_wrist": np.zeros_like(cam_wrist), |
| 72 | + }, |
| 73 | + "prompt": prompt, |
| 74 | + } |
| 75 | + |
| 76 | + |
| 77 | +def _first_action(action_result: Any) -> np.ndarray: |
| 78 | + actions = np.asarray(action_result, dtype=np.float32) |
| 79 | + if actions.ndim == 0: |
| 80 | + raise ValueError("model returned scalar action") |
| 81 | + if actions.ndim == 1: |
| 82 | + action = actions |
| 83 | + else: |
| 84 | + action = actions[0] |
| 85 | + if action.shape[0] < LEFT_ACTION_DIM: |
| 86 | + raise ValueError(f"action must have at least {LEFT_ACTION_DIM} dims, got {action.shape}") |
| 87 | + return action |
| 88 | + |
| 89 | + |
| 90 | +def action_to_move_data(action: np.ndarray, *, gripper_min: float, gripper_max: float) -> dict[str, Any]: |
| 91 | + left = np.asarray(action[:LEFT_ACTION_DIM], dtype=np.float32).reshape(-1) |
| 92 | + joint = left[:6] |
| 93 | + gripper = float(np.clip(left[6], gripper_min, gripper_max)) |
| 94 | + return { |
| 95 | + "arm": { |
| 96 | + "left_arm": { |
| 97 | + "joint": joint, |
| 98 | + "gripper": gripper, |
| 99 | + } |
| 100 | + } |
| 101 | + } |
| 102 | + |
| 103 | + |
| 104 | +def maybe_call(client: ModelClient, func_name: str) -> None: |
| 105 | + try: |
| 106 | + client.call(func_name=func_name) |
| 107 | + except Exception as exc: |
| 108 | + print(f"[panthera-client] ignore {func_name!r} RPC failure: {exc}") |
| 109 | + |
| 110 | + |
| 111 | +def run(args: argparse.Namespace) -> None: |
| 112 | + robot = PantheraSingle(enable_cameras=not args.no_cameras) |
| 113 | + client = ModelClient(host=args.host, port=args.port, timeout=args.timeout) |
| 114 | + period = 1.0 / args.hz |
| 115 | + |
| 116 | + try: |
| 117 | + robot.set_up() |
| 118 | + if not args.no_reset: |
| 119 | + robot.reset() |
| 120 | + maybe_call(client, args.reset_rpc) |
| 121 | + |
| 122 | + print("[panthera-client] ready. Press ENTER to start inference, Ctrl+C to stop.") |
| 123 | + input() |
| 124 | + print("[panthera-client] running...") |
| 125 | + |
| 126 | + step = 0 |
| 127 | + while args.max_steps <= 0 or step < args.max_steps: |
| 128 | + loop_start = time.monotonic() |
| 129 | + raw = robot.get() |
| 130 | + obs = build_observation(raw, args.prompt) |
| 131 | + action_result = client.call(func_name=args.infer_rpc, obs=obs) |
| 132 | + action = _first_action(action_result) |
| 133 | + move_data = action_to_move_data( |
| 134 | + action, |
| 135 | + gripper_min=args.gripper_min, |
| 136 | + gripper_max=args.gripper_max, |
| 137 | + ) |
| 138 | + |
| 139 | + if args.dry_run: |
| 140 | + print(f"[panthera-client] step={step} action={np.array2string(action[:7], precision=4)}") |
| 141 | + else: |
| 142 | + robot.move(move_data) |
| 143 | + |
| 144 | + step += 1 |
| 145 | + if args.log_every > 0 and step % args.log_every == 0: |
| 146 | + print(f"[panthera-client] step={step}") |
| 147 | + |
| 148 | + elapsed = time.monotonic() - loop_start |
| 149 | + sleep_time = period - elapsed |
| 150 | + if sleep_time > 0: |
| 151 | + time.sleep(sleep_time) |
| 152 | + |
| 153 | + except KeyboardInterrupt: |
| 154 | + print("\n[panthera-client] interrupted by user") |
| 155 | + finally: |
| 156 | + try: |
| 157 | + robot.shutdown(damped=args.brake_on_exit) |
| 158 | + finally: |
| 159 | + client.close() |
| 160 | + |
| 161 | + |
| 162 | +def parse_args() -> argparse.Namespace: |
| 163 | + parser = argparse.ArgumentParser(description="Panthera single-arm model-client deployment loop.") |
| 164 | + parser.add_argument("--host", default="localhost", help="model server host") |
| 165 | + parser.add_argument("--port", type=int, default=9999, help="model server TCP port") |
| 166 | + parser.add_argument("--timeout", type=float, default=30.0, help="socket timeout seconds") |
| 167 | + parser.add_argument("--infer-rpc", default="infer", help="model server inference method name") |
| 168 | + parser.add_argument("--reset-rpc", default="reset_episode", help="model server episode reset method name") |
| 169 | + parser.add_argument("--prompt", default="", help="language prompt sent with each observation") |
| 170 | + parser.add_argument("--hz", type=float, default=10.0, help="control loop frequency") |
| 171 | + parser.add_argument("--max-steps", type=int, default=0, help="0 means run until Ctrl+C") |
| 172 | + parser.add_argument("--log-every", type=int, default=10, help="print progress every N steps; <=0 disables") |
| 173 | + parser.add_argument("--gripper-min", type=float, default=0.0, help="minimum normalized gripper command") |
| 174 | + parser.add_argument("--gripper-max", type=float, default=1.0, help="maximum normalized gripper command") |
| 175 | + parser.add_argument("--dry-run", action="store_true", help="run inference but do not move the robot") |
| 176 | + parser.add_argument("--no-reset", action="store_true", help="do not reset robot before starting") |
| 177 | + parser.add_argument("--no-cameras", action="store_true", help="initialize robot without cameras") |
| 178 | + parser.add_argument("--brake-on-exit", action="store_true", help="brake instead of stop on exit") |
| 179 | + return parser.parse_args() |
| 180 | + |
| 181 | + |
| 182 | +if __name__ == "__main__": |
| 183 | + run(parse_args()) |
0 commit comments