PutCab-AB-RandomDelta-Train50

50 qualified episodes, 11,614 frames, with three 640×480 RGB cameras at 50/3 FPS. This revision replaces the earlier prefix-selected Train50 with a timing-balanced selection from the current frozen pool of 147 qualified trajectories (117 reusable v3 and 30 completed new V2). The 200-trajectory generation remains ongoing; this release is a subset of the frozen current pool, rather than a completed 200-episode parent release.

Selection

Normalized reference offset is u_ref=(delta+T_B_ref)/(T_A_ref+T_B_ref), where delta is right/B start minus left/A start. Reference durations trim exact endpoint holds and exclude added waits. Ten intervals [0,.1), …, [.9,1] each contribute five trajectories. Bins 0–4 select negative deltas; bins 5–9 select positive deltas, giving 25 negative and 25 positive episodes. Selection is without replacement within each eligible bin/sign stratum, followed by shuffling, using NumPy PCG64 seed 2026090905. It is deliberately conditioned on bin and sign, and is distinct from a uniform random sample of the entire source pool. The recorded bin balance applies to normalized reference offset; raw seconds across Seed durations form a mixture.

Per-episode offsets and distribution record the measured selection. meta/selection.json preserves eligible stratum sizes, snapshot hash, pool indices and stable source-plus-candidate parent IDs. meta/pool-snapshot.json records the frozen qualified pool. All 50 original source Seeds remain eligible in generation; this selected set may cover fewer distinct Seeds.

Expert behavior and evidence

Reused trajectories retain independent active-arm v3 timing. New V2 trajectories preserve arm paths and add the episode's recorded feedback waits: negative-delta motion lifts to a safe waypoint before approaching the cabinet and checks drawer readiness again before placement. Pose/path progression pauses during these waits while the other arm continues. Episode-specific expert_policy, reference_stable_seconds, wait_seconds, and drawer_gate_events are retained in episode metadata and lineage. Policy effects and timing balance should be considered separately when interpreting training results.

Every selected payload has verified task success and full-scene maximum penetration ≤10 mm inclusive, including robot, camera bodies and cabinet contacts. The saved hashed controls are replayed to render observations and recheck qualification. Both wrist cameras have 90° vertical FOV. All video frames, timestamps, camera calibration, path progression and payload hashes are validated. meta/qualification retains rendering, validation and retiming receipts.

observation.state contains 14 joint/gripper drive targets; action is the next observation target with the terminal target repeated. Physics runs at 250 Hz and observations every 15 ticks give 50/3 FPS. Episode/global/frame IDs are contiguous and all rows use the train split. Per-episode and full selected-set numeric/image statistics are included. Source instructions remain unchanged.

Provenance

Input trajectories: Shiki42/robotwin_putcab_fixed_leftDrawer_rightObject_AthenB_50 at 793556b0d670af13653dcb6a4ba99864fd985c59. Raw Seed/scene mapping: Shiki42/robotwin_putcab_fixed_role_source75 at b5ffd87a13cde9a376caca08245b6f202319cb00. RoboTwin base 59ccd4a347396bf30bf324a61bba9b4cdf63a2d6; immutable runtime parallelvla-2026-08-22.2. Source/policy/script identities are recorded in provenance and qualification files.

RoboTwin code carries MIT, copyright 2025 Tianxing Chen and contributors. A blanket license for upstream data and rendered third-party assets remains unspecified. SHA256SUMS binds this card and all artifact files; the local content identity is explicit in .ctr-revision.json, while the external publishing receipt records the verified immutable public revision. Evidence lifecycle: reported / audit pending. Training and policy evaluation remain separate experiments.