minimal_evaluation.py

#!/usr/bin/env python3
import os
import time
import torch

from lerobot.policies.smolvla.modeling_smolvla import SmolVLAPolicy
from lerobot.cameras.opencv.configuration_opencv import OpenCVCameraConfig
from lerobot.policies.utils import build_inference_frame, make_robot_action
from lerobot.robots.so101_follower import SO101FollowerConfig, SO101Follower
from lerobot.datasets.utils import hw_to_dataset_features
from lerobot.policies.factory import make_pre_post_processors

# -------------------------
# CONFIG
# -------------------------
MODEL_ID = "lerobot/smolvla_base"
DEVICE = "cuda" if torch.cuda.is_available() else "cpu"

FOLLOWER_PORT = "/dev/ttyACM3"
TOP_CAM_INDEX = 7
WRIST_CAM_INDEX = 9

TASK = "Pick up the red block."
ROBOT_TYPE = "so101_follower"

FPS = 10
STEPS_PER_EPISODE = 30

os.environ.setdefault("TOKENIZERS_PARALLELISM", "false")

# -------------------------
# Load model
# -------------------------
print("[INFO] Loading SmolVLA policy")
policy = SmolVLAPolicy.from_pretrained(MODEL_ID).to(DEVICE)
policy.eval()

# -------------------------
# Load OG preprocess + postprocess
# -------------------------
print("[INFO] Loading policy preprocess + postprocess")
preprocess, postprocess = make_pre_post_processors(
    policy.config,
    MODEL_ID,
    preprocessor_overrides={"device_processor": {"device": str(DEVICE)}},
)

# -------------------------
# Setup robot
# -------------------------
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),
}

robot = SO101Follower(
    SO101FollowerConfig(
        port=FOLLOWER_PORT,
        id="so101_follower_arm",
        cameras=camera_cfg,
    )
)
robot.connect()

# -------------------------
# ds_features (required glue)
# -------------------------
action_feats = hw_to_dataset_features(robot.action_features, "action")
obs_feats = hw_to_dataset_features(robot.observation_features, "observation")
ds_features = {**obs_feats, **action_feats}

# -------------------------
# Control loop
# -------------------------
print("[INFO] Starting evaluation")
policy.reset()

try:
    for step in range(STEPS_PER_EPISODE):
        obs = robot.get_observation()
        
        # ---- BUILD INFERENCE FRAME ---- 
        # save obs['camera1'] and obs['camera2'] to disk for debugging


        obs_frame = build_inference_frame(
            observation=obs,
            ds_features=ds_features,
            device=DEVICE,
            task=TASK,
            robot_type=ROBOT_TYPE,
        )

        # ---- PREPROCESS ----
        batch = preprocess(obs_frame)

        # ---- POLICY ----
        with torch.no_grad():
            raw_action = policy.select_action(batch)


        action = postprocess(raw_action)

        print("Raw action:", raw_action, "-> Decoded action:", action)
        # ---- SEND TO ROBOT ----
        
        import pdb; pdb.set_trace()
        robot_action = make_robot_action(action.squeeze(0), ds_features)
        robot.send_action(robot_action)

        time.sleep(1.0 / FPS)

except KeyboardInterrupt:
    print("\n[INFO] Ctrl+C received")

finally:
    robot.disconnect()
    print("[INFO] Done")