vibecodemaster/white-lid

White lid onto black tray

Teleoperated ElRobot demonstrations of move the white lid onto the black tray.

• LeRobot v3.0 format; 91 episodes, 23,261 timesteps, 20 FPS.
• Two RGB cameras: observation.images.camera_0 and observation.images.camera_1, each 224 × 398 pixels, H.264 video.
• Eight-dimensional observation.state and action, ordered motor_1.pos through motor_8.pos. Joint positions are calibrated to [-100,100]; gripper to [0,100]. Actions are absolute follower goals after mirroring deadband and current protection.
• One task; all episodes in the train split.

Curation and validation

The initial 59-episode release was curated as follows.

The original collection contained 61 episodes and 17,116 timesteps. Original episode 32 (four-frame accidental take) and 57 (46-frame aborted take without an object transfer) were excluded. Episode indices were compacted to 0–58; meta/curation.json records the original-to-current mapping.

Original episodes 3, 14, and 33 (current 3, 14, and 32) include human reset footage before a successful transfer; they remain included and may benefit from trimming.

All 34,232 original camera frames decoded successfully. The retained 118 videos were verified byte-for-byte against the originals; all 17,066 retained state/action rows were checked for exact preservation. Numeric values, row counts, indices, timestamps, and video dimensions were audited. Visual review used sampled timelines and endpoints across all 61 original episodes; this is not an exhaustive frame-by-frame assessment of task success.

Actual capture timing is available in meta/recording/ for 53 retained episodes. Earlier episodes only have nominal fixed-rate timestamps; those do not establish the original acquisition cadence. Cameras were sampled from the latest source frames with bounded source age/skew, rather than hardware synchronization.

October 18, 2026 update

Added 32 recorded episodes (indices 59–90), containing 6,195 timesteps. The dataset now contains 91 episodes and 23,261 timesteps. All 177 previously published data/video files remain byte-for-byte identical; the initial curation exclusions remain in effect. meta/curation.json documents the initial release only.

All 23,261 table rows and 46,522 camera frames across 182 videos passed the structural audit, with no missing or unindexed data/video files. State/action shapes, finite values, calibration bounds, indices, episode boundaries, timestamps, and video decoding were checked. This does not establish visual task success for the newly added episodes.

All recorded episodes remain available in this raw dataset revision. Subsequent sampled visual review recommends excluding 11 episodes from clean training; see below. Episodes 73 and 86 have joint steps above 20 normalized units per frame during retreat; sampled video shows successful placements, and the sensor warning is retained. Actual capture timing sidecars are available for 53 of 91 episodes. meta/validation/2026-10-18.json contains per-episode audit results; the accompanying manifest records SHA-256 hashes of the uploaded dataset files.

Sampled visual review and clean training selection

A parallel review examined both cameras at eight evenly spaced timestamps across all 91 episodes, supported by numeric audits and targeted checks of flagged joint discontinuities. This is sampled visual inspection, not exhaustive frame-level verification or a standardized robot evaluation.

For conservative clean training, exclude episodes 3, 14, 32, 60, 62, 63, 64, 65, 66, 67, 68. Episodes 3, 14, 32, and 60 contain human reset prefixes; 63–68 are human reset clips with stationary or nearly stationary robot actions. Episode 62 reaches the tray but completed release cannot be established at its endpoint. The untrimmed raw recordings are preserved. This selection retains 80 episodes / 21,538 timesteps, including 24 new episodes / 5,428 timesteps.

meta/training_curation_2026-10-18.json provides exact episode lists, exclusions, denominators, and suggested shared train/evaluation splits. These splits hold out original episodes 53–58 and new episodes 87–90 (10 episodes / 2,324 timesteps). The remaining combined training set has 70 episodes / 19,214 timesteps; the new-data-only training set has 20 episodes / 4,648 timesteps. No held-out episode contributes optimizer training samples. The historical 4k checkpoint was trained before these new exclusions; continuing it does not erase that training history.

The structural audit and hash manifest describe the frozen source snapshot at revision 00fc1a46721b043ebaf2ed23776a82a5de68401a, before this documentation-only review update.

Usage

from lerobot.datasets.lerobot_dataset import LeRobotDataset
dataset = LeRobotDataset("vibecodemaster/white-lid")

Usable as a LeRobot training dataset with a compatible robot/policy configuration. Model-specific action, observation, and normalization adapters may be needed for other frameworks.