Shiki42/PutCab-AB-RandomDelta-Train50-V3

PutCab AB RandomDelta Train50 V3 (E098)

50 physically qualified demonstrations from 50 distinct source seeds in LeRobot v2.1 format. All three camera streams are 320×240 H.264 at 50/3 FPS. There are 11,903 frames and 150 videos.

Open in the official LeRobot Dataset Visualizer

Enter Shiki42/PutCab-AB-RandomDelta-Train50-V3 to review episodes 0–49.

Sampling and qualification

Normalized delta is u=(delta+T_B_ref)/(T_A_ref+T_B_ref). Every one of 50 equal strata of [0,1) contains exactly one episode, before and after 4 ms quantization. Ten equal bins each contain five episodes. Seconds delta spans -9.432 to +6.692 seconds (29 negative, 21 positive). Uniformity applies to normalized delta; each source trajectory supplies its own reference durations.

45 original qualified source identities are retained. Five supplemental identities were uniformly drawn from the remaining serial-qualified sources of the same revision-pinned source75 pool. Every original sampled u is retained. Supplemental seeds: 7810, 9410, 7016, 5643, 10609. Source selection is conditioned on task and physics qualification.

Positive offsets retain independent active-arm timelines. Negative offsets retain two-stage drawer waiting gates, with realized waits recorded separately. R003 onward uses a 0.14 m drawer readiness threshold and 0.3 s stability. All delivered episodes pass task success and maximum scene penetration at most 10 mm; physics runs at 250 Hz.

31 retained R001 episodes were resized using Lanczos, explicit 50/3 FPS timestamps and scaled camera intrinsics. The remaining 19 were rendered natively at 320×240. Camera extrinsics and state/action payloads retain their validated identities.

Contents

• data/ and videos/: LeRobot per-episode numerical data and three camera videos.
• meta/info.json, episodes.jsonl, tasks.jsonl, and statistics: LeRobot metadata.
• meta/episode_delta.csv, lineage.jsonl, and final-source-mapping.json: per-episode seed and delta provenance.
• meta/norm_stats.json: pinned OpenPI normalization over all 11,903 frames, action horizon 50.
• meta/norm-stats-receipt.json: normalization source and SHA-256 identity.
• meta/calibration/: per-frame camera calibration.
• SHA256SUMS: publication manifest, including this card and the source identity marker.

Immutable source identity

Raw source75: Shiki42/robotwin_putcab_fixed_role_source75, revision b5ffd87a13cde9a376caca08245b6f202319cb00.

Verified local dataset manifest SHA-256: e48786ea7367d058b670b9b3ec53e5bec66b2f47e2f9eaa4d7afb71e11a4291d. Normalization SHA-256: bd8c65bdbe61b78d7ccaf953c92d17a8e4ec5e5e4fb556007bf739407abe06ef.

.ctr-revision.json preserves the verified local source snapshot identity. meta/source-SHA256SUMS preserves its original manifest; its README entry refers to meta/source-README.md. The root publication manifest and Hugging Face commit identify the published tree. Data, videos, original numerical metadata and normalization are byte-identical to that verified source.

Review status

50/50 delivered episodes passed physical qualification and CPU frame/media verification. Result lifecycle: reported → audit pending. User visual review and audit confirmation remain pending. Training and checkpoint evaluation are separate experiments.