PutCab-AB-RandomDelta-V2-200

200 qualified episodes, 46,725 frames, three640×480 camera streams at50/3FPS. Complete normalized-reference-bin-balanced200 release; previously published balanced Train50 stays unchanged.

Open official LeRobot Dataset Visualizer

V2 scope and distribution

The complete V2 release reuses 117 previously qualified trajectories and adds 83 deficit-directed qualified trajectories. It contains exactly200 unique episodes, with18–22 per normalized timing bin. The independently published balanced Train50 remains at its existing revision. Existing qualified episodes outside quotas are retained in the external donor pool; source payload hashes and parent indices are preserved in meta/lineage.jsonl.

Delta is right/B start minus left/A start in seconds. The reference coordinate is u_ref=(delta+T_B_ref)/(T_A_ref+T_B_ref), using each original Seed's fixed v3 active-arm durations. Reference durations trim exact initial/final holds. Positive delta delays the right arm; negative delta delays the left. All50 original solvable Seeds remain eligible. Raw seconds across different Seed durations form a mixture, so the quota target applies to normalized offsets. Realized insertion time includes separately recorded waits.

New proposals target only deficient bins, choose a source Seed uniformly and sample continuously within the chosen reference bin (PCG64 seed2026090902), then round to4ms ticks. Final collection retains candidate order independently of worker completion order. Reuse selection seed2026090901; complete-dataset shuffle seed2026090904. Qualification changes the within-bin distribution; the claimed balance is the measured bin-count condition.

Normalized reference interval	Episodes
0.0–0.1	18
0.1–0.2	18
0.2–0.3	18
0.3–0.4	18
0.4–0.5	18
0.5–0.6	22
0.6–0.7	22
0.7–0.8	22
0.8–0.9	22
0.9–1.0	22

This release contains 103 negative and 97 positive sampled deltas. Detailed distribution · Per-episode delta CSV. meta/episodes.jsonl records sampled/applied delta, normalized coordinate, bin, source Seed, policy version, reused/new provenance and wait durations for each episode.

Negative-delta expert policy

V2 adds conditional control, not merely a new starting offset. The right arm grasps and lifts the object to a stable safe waypoint before approaching the cabinet. It holds the gripper closed until the actual drawer target has translated at least0.15m and stayed ready for the episode-specific reference stability interval (0.1s for early candidates and0.3s for later candidates, recorded as reference_stable_seconds). The right arm then approaches its existing above-target pre-placement plateau; readiness is checked again before descent and release. Source arm paths retain their original changing commands and internal60ms progress intervals; waiting pauses source-path progression while the other arm continues. Timeout, contact or task failures are rejected.

The expert's feedback reference pass generates and saves the complete retimed controls and gate events. A separate replay loads that exact hashed control file for rendering and rechecks task success, contacts and actual drawer opening at recorded release events. This second replay, rather than the planning pass alone, establishes final qualification. Complete physical-tick progress maps are validated against both original arm trajectories; metadata marks negative_two_stage_drawer_gate_v2. Reused data retains its v3 policy identity. Report policy-specific effects when comparing models: V2 changes both data balance and negative-timing expert behavior.

Quality and training interface

All accepted trajectories pass task success and full-scene maximum penetration≤10mm at every4ms tick, including camera bodies and cabinet contacts. Both wrist cameras use90° vertical FOV. Video frame count/timestamps, camera transforms, state/action alignment and file hashes are checked. observation.state is the14 joint/gripper drive target vector; action is the next observation target, terminal target repeated. Episode/global/frame indices are contiguous; all rows use the train split. Per-episode and aggregate statistics and camera calibration are included. Physics is250Hz; observations every15ticks yield50/3FPS.

Source identities and licensing

Original source: Shiki42/robotwin_putcab_fixed_leftDrawer_rightObject_AthenB_50 at revision 793556b0d670af13653dcb6a4ba99864fd985c59. Raw scene mapping: Shiki42/robotwin_putcab_fixed_role_source75 at b5ffd87a13cde9a376caca08245b6f202319cb00. RoboTwin base commit 59ccd4a347396bf30bf324a61bba9b4cdf63a2d6; runtime parallelvla-2026-08-22.2. Source evidence, scripts and qualification hashes are recorded in meta/provenance.json and meta/qualification.

RoboTwin code is MIT, copyright2025 Tianxing Chen and contributors. The upstream data/assets expose no blanket derived-data license; that licensing remains unspecified. SHA256SUMS covers the payload/card, .ctr-revision.json identifies local content, and external publication receipts record immutable public HF revisions. Accelerated continuation uses v2_validate_fast with exact uint8 histogram statistics while preserving every decoded-frame, timestamp, camera, source-path and hash check. Earlier validation receipts retain their original identity. Dataset lifecycle: reported / audit pending. Training and policy evaluation remain separate experiments.