2sin0/deux-pick-scan-place-46

DEUX pick–scan–place: clean

Direct successes, with no recorded failed attempt or scan search.

This dataset contains 46 episodes, 49,209 aligned frames, and 138 camera videos at 30 FPS, in LeRobotDataset v2.1 format. It is part of a 50-success collection (46 direct successes and 4 recovered successes). The companion dataset is 2sin0/deux-pick-scan-place-recovered-4.

The right hand physically picks a bag from moving source boxes, the left-hand fixed reader scans the barcode, and the right hand places the bag in a collection box in MuJoCo. Object pose is read from simulator ground truth. Grasp assistance is disabled. Barcode success uses simulated geometric alignment and dwell conditions, not image decoding.

Capture and variation

• Collection seeds: 7–56 across the full collection; exact per-episode seed and raw ID are in meta/source_episodes.json.
• Exact grasp, presentation, and release endpoints; intermediate transport waypoint variation bounded by 3 mm.
• Per-episode free-space speed scale sampled in [0.95, 1.0]; acceleration limits scale quadratically.
• Source control period: 34 ms. Export: causal previous-sample alignment at 30 FPS, maximum source age/gap 40 ms.
• Some source frames are reused when aligning the approximately 29.41 Hz control stream to 30 FPS.
• Each saved episode has a matching successful scan and placement within the same reservation.

Features

Feature	Shape / type	Meaning
observation.state	float32[30]	Measured end-effector poses, arm joints, grippers
action	float32[30]	Absolute joint/gripper targets and FK of those arm targets
observation.images.head	RGB 360×640, MP4	Head camera
observation.images.left_wrist	RGB 360×640, MP4	Left wrist camera
observation.images.right_wrist	RGB 360×640, MP4	Right wrist camera
annotation.phase	int64	Main phase ID
annotation.subphase	int64	Fine-grained phase ID
annotation.recovery	int64	1 from the first failure/search event through episode end

The 30-D vector order is left EE pose (7), right EE pose (7), left arm joints (7), right arm joints (7), left gripper (1), right gripper (1). EE pose is x, y, z, qx, qy, qz, qw in world coordinates (meters, unit quaternion). Joints are radians; grippers are normalized with 0 closed and 1 open. Names are specified in meta/info.json.

Labels

Phase ID	Label
0	idle
1	pick
2	scan
3	place
4	recovery

Subphase IDs: 0: idle, 1: park, 2: approach, 3: wait_arrival, 4: descend, 5: grasp, 6: lift, 7: scan_approach, 8: scan_align, 9: scan_search, 10: scan_return, 11: transport, 12: lower, 13: release, 14: return, 15: recover.

Normal return after placement is place / return, not failed recovery. annotation.recovery is a persistent failure-history flag; phase=recovery labels actual failed-attempt arm recovery. The label dictionaries are in meta/label_maps.json. Detailed stage, failure, scan, and placement events and controller summaries are in alignment/episode_*/episode_context.json. Source timestamps and alignment indices are preserved.

Loading and compatibility

This is v2.1, validated with the upstream LeRobot v0.3.3 loader. It is not a v3.0 dataset. Use a v2.1-compatible LeRobot installation, then:

from lerobot.datasets.lerobot_dataset import LeRobotDataset

dataset = LeRobotDataset("2sin0/deux-pick-scan-place-46", revision="v2.1", video_backend="pyav")
sample = dataset[0]

Validation and provenance

All 50 source episodes passed full raw stream integrity and synchronized-label checks. The converter verifies every MP4 frame's count, dimensions, FPS, and timestamp, and checks raw source SHA-256 before and after conversion. Every assembled Parquet feature other than remapped episode/global indices was compared with its independently exported source. The upstream loader reads first/middle/last frames of every episode, all three cameras, and checks delta-timestamp queries and boundary padding. See meta/validation.json.

meta/conversion.json contains source hashes and conversion settings; meta/source_episodes.json maps exported indices to original attempts and seeds. Statistics use all aligned numeric rows and pre-encode RGB pixels (MP4 uses lossy H.264). All episodes are in the train split; no held-out evaluation split is defined.