wenyixu101/so101-pick-and-place

SO-101 Teleoperation Pick-and-Place

SO-101 pick-and-place demonstrations collected by human teleoperation with a physical SO-101 leader and follower arm.

• Dataset: wenyixu101/so101-pick-and-place
• Dataset version: 0.0.0

Data provenance

Dimension	Value
Environment	Real-world tabletop workspace
Action source	Human teleoperation with a physical SO-101 leader arm
Robot embodiment	Physical SO-101 follower arm

Dataset summary

Metric	Value
Task instruction	Pick up the object and place it on the target.
Robot	SO-101 follower arm: 5 arm joints + 1 gripper joint
Control source	Physical SO-101 leader arm
Episodes	50 demonstrations
Frames	18,250 at 25 FPS (12 min 10 sec)
Split	Train: episodes 0–49
Top camera	RGB video, 1280 × 720 at 25 FPS
Wrist camera	RGB video, 640 × 480 at 25 FPS
Format	LeRobotDataset v3.0

Episode preview

Top camera

Wrist camera

Collection setup and variation

The demonstrations use one object type and one target type. The initial object and target positions are varied between episodes while the tabletop setup and task semantics remain consistent. The external top-camera placement may have small pose differences, but no large viewpoint changes were intentionally introduced.

The dataset contains successful human demonstrations retained after manual episode review. It does not provide a formal held-out validation or test split.

Policy features

Feature	Shape / format	Description
observation.state	float32[6]	Current SO-101 joint positions
observation.images.top	video, 720 × 1280 × 3	External top-camera RGB
observation.images.wrist	video, 480 × 640 × 3	Wrist-mounted camera RGB
action	float32[6]	Teleoperated joint-position command

Joint order for observation.state and action: shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, and gripper.

Intended use

This dataset is intended for robot-learning research, behavior cloning, LeRobot data-pipeline development, and physical SO-101 pick-and-place experiments. Report the immutable dataset tag used for training or evaluation.

The data covers a single object type, a single target type, one robot setup, and a limited collection session. Performance obtained from this dataset should not be interpreted as evidence of broad generalization to unseen objects, targets, workspaces, viewpoints, or robot hardware.

Validation performed before publication

• 50 Parquet files and 18,250 rows were present.
• Episode indices were contiguous from 0 through 49.
• Frame indices were contiguous within every episode.
• 50 top-camera and 50 wrist-camera videos were present.
• Every video was fully decoded; decoded frame counts matched the corresponding episode data for both cameras.

License

The dataset is released under the Creative Commons Attribution 4.0 International license (CC BY 4.0): https://creativecommons.org/licenses/by/4.0/

Versioning and attribution

Dataset version 0.0.0 is the initial release containing 50 physical-robot human teleoperation demonstrations. Cite this dataset repository and the immutable version tag when using it.