#!/usr/bin/env python3 import time import sys import queue import inspect import torch from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig from lerobot.datasets.utils import hw_to_dataset_features from lerobot.policies.factory import make_pre_post_processors from lerobot.policies.utils import build_inference_frame, make_robot_action from lerobot.robots.so101_follower import SO101FollowerConfig, SO101Follower from lerobot.teleoperators.so101_leader import SO101LeaderConfig from lerobot.teleoperators import make_teleoperator_from_config from lerobot.datasets.lerobot_dataset import LeRobotDataset # ------------------------- # CONFIG # ------------------------- FOLLOWER_PORT = "/dev/ttyACM1" LEADER_PORT = "/dev/ttyACM2" TOP_CAM_INDEX = 4 WRIST_CAM_INDEX = 9 MODEL_ID = "lerobot/smolvla_base" TASK = "Pick up the red block." ROBOT_TYPE = "so101_follower" FPS = 20 POLICY_SCALE = 1 EPISODE_SECONDS = 10.0 # ---- Recording / Hub ---- curr_time = time.strftime("%Y%m%d_%H%M%S", time.localtime()) DATASET_REPO_ID = f"HenryZhang/so101_smolvla_eval_{curr_time}" DATASET_ROOT = None USE_VIDEOS = True PUSH_TO_HUB_ON_EXIT = True PRIVATE_ON_HUB = False DATASET_TAGS = ["LeRobot", "so101", "smolvla", "policy-eval"] # ------------------------- def log(msg): print(msg, flush=True) def start_enter_listener(cmd_q: "queue.Queue[str]"): """Press Enter to start one episode.""" while True: try: sys.stdin.readline() cmd_q.put("start_episode") except Exception: break def send_leader_action(robot, leader_action): if not isinstance(leader_action, dict): return out = {k: float(leader_action[k]) for k in robot.action_features.keys() if k in leader_action} if out: robot.send_action(out) def _import_build_dataset_frame(): try: from lerobot.common.datasets.utils import build_dataset_frame return build_dataset_frame except Exception: from lerobot.datasets.utils import build_dataset_frame return build_dataset_frame def create_dataset(repo_id, fps, root, robot_type, features, use_videos, num_cameras): kwargs = dict( repo_id=repo_id, fps=fps, root=root, robot_type=robot_type, features=features, use_videos=use_videos, image_writer_processes=0, image_writer_threads=4 * max(num_cameras, 1), ) try: if "single_task" in inspect.signature(LeRobotDataset.create).parameters: kwargs["single_task"] = TASK except Exception: pass try: ds = LeRobotDataset.create(**kwargs, exist_ok=True) except TypeError: ds = LeRobotDataset.create(**kwargs) if hasattr(ds, "start_image_writer") and num_cameras > 0: ds.start_image_writer(num_processes=0, num_threads=4 * num_cameras) log(f"[INFO] Dataset ready: {repo_id}") return ds def dataset_add_frame_compat(dataset, frame, task): try: if "task" in inspect.signature(dataset.add_frame).parameters: dataset.add_frame(frame, task=task) return except Exception: pass frame["task"] = task dataset.add_frame(frame) def dataset_push_compat(dataset, repo_id, tags, private): try: if len(inspect.signature(dataset.push_to_hub).parameters) >= 1: dataset.push_to_hub(repo_id, tags=tags, private=private) return except Exception: pass dataset.push_to_hub(tags=tags, private=private) def main(): device = torch.device("cuda" if torch.cuda.is_available() else "cpu") log(f"[INFO] Device: {device}") build_dataset_frame = _import_build_dataset_frame() # ---- Load policy ---- log(f"[INFO] Loading SmolVLA: {MODEL_ID}") policy = SmolVLAPolicy.from_pretrained(MODEL_ID).to(device) policy.eval() preprocess, postprocess = make_pre_post_processors( policy.config, MODEL_ID, preprocessor_overrides={"device_processor": {"device": str(device)}}, postprocessor_overrides={"device_processor": {"device": str(device)}}, ) # ---- Cameras ---- camera_cfg = { "camera1": OpenCVCameraConfig(index_or_path=TOP_CAM_INDEX, width=640, height=480, fps=30), "camera2": OpenCVCameraConfig(index_or_path=WRIST_CAM_INDEX, width=640, height=480, fps=30), } # ---- Robots ---- robot_cfg = SO101FollowerConfig(port=FOLLOWER_PORT, id="so101_follower_arm", cameras=camera_cfg) leader_cfg = SO101LeaderConfig(port=LEADER_PORT, id="so101_leader_arm") log("[INFO] Connecting follower...") robot = SO101Follower(robot_cfg) robot.connect() log("[INFO] Connecting leader...") teleop = make_teleoperator_from_config(leader_cfg) teleop.connect() # ---- Dataset ---- action_features = hw_to_dataset_features(robot.action_features, "action", USE_VIDEOS) obs_features = hw_to_dataset_features(robot.observation_features, "observation", USE_VIDEOS) dataset_features = {**action_features, **obs_features} dataset = create_dataset( DATASET_REPO_ID, FPS, DATASET_ROOT, robot.name, dataset_features, USE_VIDEOS, len(getattr(robot, "cameras", [])), ) # ---- Enter listener ---- cmd_q = queue.Queue() import threading threading.Thread(target=start_enter_listener, args=(cmd_q,), daemon=True).start() log("\n[INFO] Press Enter to run ONE episode. Ctrl+C to exit.\n") dt = 1.0 / FPS mode = "RESET" episode_idx = 0 episode_end_time = None policy.reset() try: while True: t0 = time.time() if mode == "RESET" and not cmd_q.empty(): cmd_q.get_nowait() episode_idx += 1 policy.reset() if hasattr(dataset, "clear_episode_buffer"): dataset.clear_episode_buffer() episode_end_time = time.time() + EPISODE_SECONDS mode = "POLICY" log(f"[INFO] Episode {episode_idx} START") if mode == "RESET": send_leader_action(robot, teleop.get_action()) else: if time.time() >= episode_end_time: log(f"[INFO] Episode {episode_idx} END — saving...") t_save = time.time() dataset.save_episode() log(f"[INFO] Saved in {time.time() - t_save:.1f}s") mode = "RESET" episode_end_time = None else: obs = robot.get_observation() obs_frame = build_inference_frame( observation=obs, ds_features=dataset_features, device=device, task=TASK, robot_type=ROBOT_TYPE, ) with torch.no_grad(): batch = preprocess(obs_frame) action = policy.select_action(batch) action = postprocess(action) if isinstance(action, torch.Tensor): action = action.squeeze(0) * POLICY_SCALE robot_action = make_robot_action(action, dataset_features) sent_action = robot.send_action(robot_action) print("Predicted:", action, "robot:", robot_action, "sent:", sent_action) frame = { **build_dataset_frame(dataset.features, obs, "observation"), **build_dataset_frame(dataset.features, sent_action, "action"), } dataset_add_frame_compat(dataset, frame, TASK) time.sleep(max(0.0, dt - (time.time() - t0))) except KeyboardInterrupt: log("\n[INFO] Ctrl+C received.") finally: teleop.disconnect() robot.disconnect() if PUSH_TO_HUB_ON_EXIT: log("[INFO] Pushing dataset to Hub...") dataset_push_compat(dataset, DATASET_REPO_ID, DATASET_TAGS, PRIVATE_ON_HUB) log("[INFO] Done.") if __name__ == "__main__": main()