mohsinarf/so101_pick_blue_cube_into_tray

SO-ARM101 — Pick Blue Cube into Blue Tray

A LeRobot imitation-learning dataset of an SO-ARM101 6-DOF robot arm performing a pick-and-place task: grasp a blue cube and place it into a blue tray. Recorded via teleoperation with synchronized dual-camera video and joint states/actions.

Collected as part of the Intel collaboration Physical AI Studio project.

Task

"pick up the blue cube and place it in the blue tray"

Each episode is a single demonstration of the full pick-and-place motion.

Dataset at a glance

Robot	SO-ARM101 (6-DOF arm + gripper)
Format	LeRobot v3.0
Episodes	40
Frames	12,840
Recording rate	30 fps
Total duration	~7.1 min (428 s)
Episode length	249–426 frames (mean 321, ~10.7 s)
Tasks	1 (pick up and place)
Cameras	2 (fixed camera, arm camera)
Video	800×600, AV1, yuv420p, 30 fps, no audio
Split	train (all 40 episodes)

Features

Proprioception — both action and observation.state are 6-D vectors of joint positions (degrees), in this order:

shoulder_pan.pos, shoulder_lift.pos, elbow_flex.pos,
wrist_flex.pos, wrist_roll.pos, gripper.pos

• action — commanded next joint pose (teleoperator target)
• observation.state — measured current joint pose
• observation.images.fixed camera — external scene view (800×600)
• observation.images.arm camera — wrist/gripper view (800×600)
• timestamp, frame_index, episode_index, task_index, index — indexing

Structure

├── data/chunk-000/            # per-frame joint states & actions (Parquet)
├── videos/
│   ├── observation.images.fixed camera/chunk-000/
│   └── observation.images.arm camera/chunk-000/   # AV1 MP4, episodes packed per file
└── meta/
    ├── info.json              # schema, feature specs, fps, splits
    ├── tasks.parquet          # task definitions
    ├── stats.json             # global normalization statistics
    └── episodes/chunk-000/    # per-episode index (row/timestamp ranges) + stats

Multiple episodes are concatenated into shared data/video files; each episode's dataset_from_index/dataset_to_index (Parquet rows) and from_timestamp/to_timestamp (video segment) in the episode metadata let loaders reconstruct individual episodes.

Usage

Install LeRobot and load the dataset directly:

from lerobot.common.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("mohsinarf/so101_pick_blue_cube_into_tray")

print(dataset.num_episodes, dataset.num_frames)
sample = dataset[0]
print(sample["observation.state"].shape)   # (6,)
print(sample["action"].shape)              # (6,)

The dataset is compatible with LeRobot policies such as ACT and Diffusion Policy.

Citation

If you use this dataset, please cite it:

@misc{so101_pick_blue_cube_into_tray_2026,
  title  = {SO-ARM101 — Pick Blue Cube into Blue Tray},
  author = {mohsinarf},
  year   = {2026},
  note   = {LeRobot dataset, Physical AI Studio (Intel collaboration)},
  url    = {https://huggingface.co/datasets/mohsinarf/so101_pick_blue_cube_into_tray}
}

License

Released under the Apache 2.0 license.