NOJIMA21/xlerobot-final-v1-xlerobot_mobile_straight_wheelstate_v1
XLeRobot Final v1 whole-body wheel-feedback integration record A public, real-hardware LeRobot Dataset v3 integration record for an XLeRobot-compatible 0.4.0 differential two-wheel configuration. This is a transparent engineering dataset, not a household-task benchmark or an autonomy-performance claim. Recorded system Two Seeed Studio Pro SO-101 arms under leader/follower teleoperation Differential two-wheel base Pan/tilt head Three RGB cameras: left wrist… See the full description on the dataset page: https://huggingface.co/datasets/NOJIMA21/xlerobot-final-v1-xlerobot_mobile_straight_wheelstate_v1.
XLeRobot Final v1 whole-body wheel-feedback integration record
A public, real-hardware LeRobot Dataset v3 integration record for an XLeRobot-compatible 0.4.0 differential two-wheel configuration. This is a transparent engineering dataset, not a household-task benchmark or an autonomy-performance claim.
Recorded system
- Two Seeed Studio Pro SO-101 arms under leader/follower teleoperation
- Differential two-wheel base
- Pan/tilt head
- Three RGB cameras: left wrist 640x480, overhead 1280x720, right wrist 640x480
- 30 Hz H.264 video recording
Dataset schema
One 89.37-second manually teleoperated episode (2,682 frames) records a 16-dimensional observation state and matching 16-dimensional action:
- Left arm: 6 joint positions
- Right arm: 6 joint positions
- Head pan/tilt: 2 positions
- Base left/right wheel: 2 signed velocities
Wheel entries in observation.state are measured STS3215 Present_Velocity feedback; wheel entries in action are commanded Goal_Velocity. They are deliberately retained as separate signals.
Scope and limitations
The episode's natural-language task is TEST. It validates simultaneous whole-body teleoperation, three-camera H.264 recording, and the published wheel-feedback schema. It is intended as a reproducible reference for LeRobot-compatible whole-body data layout, not as training data for a manipulation or navigation success claim.
Known camera-label correction
In this initial episode only, the physical left and right wrist cameras were mounted opposite to their feature labels: observation.images.left_wrist_camera contains the physical right-wrist view, and vice versa. This affects only the two wrist-image feature names; arm state/action, head, wheel state/action, timestamps, and overhead video are unaffected.
The camera mounts were physically corrected after this integration capture. All Final v1 datasets recorded after this repository use the intended mapping: left_wrist_camera is the physical left wrist and right_wrist_camera is the physical right wrist. Users of this specific smoke-test episode should swap the two wrist-image keys at load time if physical-side identity matters.
