up-down_20261023_213122

Early return-to-start recording — not a sorting demonstration.

Part of a quality-control pipeline built at the Robotekin LeRobot Hackathon 2026: an SO-101 arm picks a metal part, a vision model checks whether it has a pin, and a sorting policy drops it in the green box (with pin) or the white box (without pin). This dataset belongs to Warm-up recording (before the final task definition).

Naming note: good / bad describe the part, not the quality of the demonstration. GOOD = part with a pin → green box. BAD = part without a pin (defective) → white box. up-down-bad_* datasets are successful demonstrations of carrying a defective part to the white box. No failed or negative demonstrations were intentionally recorded. Other informal names: so101-pick-cube_* contains the metal bracket, not a cube, and up-down_* is the GOOD-box sorting task (except the 23 October sessions, see below).

What this dataset contains

Recorded on 23 October, before the sorting task existed. Despite the up-down name and the stored task string "Move the robot up and down", the episodes are the arm's return trip to the start pose; they do not carry a part to a box. These sessions have no meta/recording_notes.json, are left out of the GOOD-box training data, and are not related to the GOOD/BAD classification.

Stored task	"Move the robot up and down"
Episodes	6 (367 frames, 36.7 s in total)
Recorded	2026-10-23 21:31
Frame rate	10 fps
Cameras	observation.images.front, observation.images.wrist (1280x720 RGB)
State / action	observation.state / action, 6 values each
Object locations	not recorded (no meta/recording_notes.json)

How it was used

• Training data of the first experiment, pick_v0_mixed in MikelRubi/act-pick (pick and return trips mixed in one policy, superseded).

Pipeline

tray  ──► act-pick ──► part held up to the camera ──► pin-classifier ──┬─ pin    (GOOD) ──► act-up-down-good ──► green box
                                                                      └─ no pin (BAD)  ──► act-up-down-bad  ──► white box

Related models: MikelRubi/act-pick → MikelRubi/pin-classifier → MikelRubi/act-up-down-good | MikelRubi/act-up-down-bad. Project page: MikelRubi/robotekin-so101-pin-sorting. On the robot the stages are chained by the team's demo app (30_full_demo_app.py); during recording the operator chose the box after looking at the part. The box for each result is set by the operator's order; the usual order is shown above.

Recording setup

• Robot: SO-101 follower arm (LeRobot so_follower, 6 joints including the gripper), teleoperated with an SO-101 leader arm (leader/follower). The policies never see the leader; action is the leader position sent to the follower.
• Cameras: two USB RGB cameras, front (fixed overhead camera looking down at the tray) and wrist (mounted on the gripper).
• Joints (state and action, same order): shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper. Arm joints in degrees, gripper 0–100 (LeRobot SO-101 calibration).
• Format: LeRobot dataset v3.0, H.264 video. Recorded on 23–24 October 2026 at the Robotekin LeRobot Hackathon (Robotory, Vitoria-Gasteiz).

Known limitations

• A single short recording session in one room and lighting setup. No success labels are stored: every saved episode is treated as a successful demonstration.
• Camera models and their exact mounting positions are not documented in the dataset.
• Recorded at 10 fps and 1280x720, unlike the later 30 fps / 640x480 sessions; training scripts up-sample and crop it to combine both.

Acknowledgements

Recorded by Iker Azurmendi, Mikel Zulueta, Ander Autillo and Diego Suarez during the Robotekin LeRobot Hackathon 2026 (23–25 October 2026, Vitoria-Gasteiz). Thanks to Bravo Robots and to Robotekin for the hackathon, to Hugging Face for LeRobot, and to everyone who took part in it.