You need to agree to share your Hugging Face information to access this dataset

This repository is publicly accessible, but you have to accept the conditions to access its files and content.

Log in or Sign Up to review the conditions and access this dataset content.

rotary-table-walnuts

LeRobotDataset v3.0 of an OpenYAM (Anvil workcell) robot teleoperated with a Quest 3S at an onsite walnut packing line (session duche-nut-onsite, recorded 2026-11-11) to "pack walnut bags from the rotary table into the box" (single task).

The task prompt was written from the video during processing; no prompt was recorded with the data.

robot_type anvil_yam
episodes / frames 2 / 24,939 (6.9 min)
fps 60
cameras chest (head), wrist_l, wrist_r — 480x270, 180° fisheye, h264
observation.state / action 14 = per arm [finger_joint1, joint1..joint6], left then right; absolute joint positions

Single-arm data. The left arm does the task (gripper used in both episodes); the right arm stays nearly parked and its gripper is not used. Its state and action columns are present but nearly constant.

Wrist camera labels look swapped. The image stream recorded as /cam_wrist_r (here observation.images.wrist_r) is the one that moves, grasps and carries the bags, so it appears to be mounted on the left (active) arm, and wrist_l on the parked right arm. Names are kept exactly as recorded so the dataset matches the live robot's topics at inference time. Check the physical camera wiring before relying on the names.

Episodes

Whole episodes, no trimming or splitting (teleop pauses are kept as recorded).

episode_index Raw episode Length Note
0 0002 215.0 s idle pauses up to 18.4 s (left arm), kept
1 0005 200.6 s idle pauses up to 40.8 s (right arm), kept

Raw episodes not included (all remain in the raw capture):

  • 0001 (85 s): mcap-valid critical — right-arm command topic absent for the whole episode.
  • 0003 (2 s), 0004 (32 s): no teleop commands at all, arms static.
  • 0006 (248 s): mcap-valid critical — wrist_l camera stopped 109 s in (no frames for the last 140 s; USB dropout).

Provenance

Raw MCAP (6 episodes, 36.1 GB): Anvil workcell ~/anvil-loader/data/recordings/duche-nut-onsite, copied to field SSD TITUS-FLD-B and checksum-verified (2026-11-15); R2 titus-datasets/captures/ upload in progress at the time of writing. Converted with anvil-embodied-ai mcap-convert (commit 5d1c8eb, config conversion_config.yaml / openyam_bimanual_quest_16x9).

QC

  • mcap-valid: both included episodes warning (idle gaps only), 0 critical. Full 6-episode report in anvil_qc/mcap_valid_all6.log.
  • dataset-valid: passed. All 14 action columns track their matching joints; no NaNs. Debug plots in anvil_qc/debug_plots/.
  • Cameras: 100% coverage, ~60 Hz, max frame gap 36 ms in both included episodes.
  • Image sharpness (median Laplacian variance, raw 1080p): chest ~880, wrist_l ~860–1,020, wrist_r ~340–450. wrist_r (the active arm's camera) is noticeably softer than the others.
  • Per-episode raw QC JSON and summary in anvil_qc/qc_raw/; frame sheets in anvil_qc/frames/.
Downloads last month
25