Albertvh97/conveyor-sort-sim-demos

Conveyor sort: SO-101 MuJoCo demonstrations

50 recorded LeRobot v3 episodes (62,866 frames at 30 FPS) from Physical AI Studio's virtual-leader teleoperation. A scripted MuJoCo demonstrator drove the virtual leader; Studio recorded the follower's camera observations, joint state, and actions. The demonstrator uses privileged simulation state, but that state is not a dataset observation.

• Scene: conveyor_sort (SO-101 follower)
• Features: six-dimensional observation.state and action, and 640×480 observation.images.overview and observation.images.wrist videos.
• All 50 episodes were recorded with physicalai-mujoco-plugin at 0.04 m/s belt speed, 20% slower than the original 0.05 m/s recording. The faster episode was removed and replaced. See simulation.json for the simulator version.
• All episodes were retained, including imperfect ones. There is no held-out evaluation split; select one explicitly before measuring policy quality. No training performance claim is made.

Load for training:

from physicalai.data import LeRobotDataModule

data = LeRobotDataModule(
    repo_id="Albertvh97/conveyor-sort-sim-demos",
    train_batch_size=2,
    num_workers=0,
)

For reproducible training, pass revision="<commit SHA>" to LeRobotDataModule. To run the Studio simulation E2E test, download the repository at that revision into a directory containing meta/, data/, videos/, and simulation.json and set SIM_E2E_DATASET_DIR to that directory. During evaluation disable the simulator autopilot: otherwise its actions can be mistaken for policy performance.