Shiki42/PutCab-Mixed-Train50-V4

PutCab Mixed V4 Train50

50 qualified demonstrations from a common 100-scene protocol. Three 320×240 H.264 cameras at 50/3 FPS, 250 Hz physical control, 14653 observations. Construction/packaging Run: E108-R004.

The left arm opens the drawer and the right arm grasps, lifts, transfers and releases the object. Sequential runs the two arm programs serially, with half of each order. Concurrent starts both together. CTR keeps frozen normalized Delta choices and records the exact signed start offset. Native planning occurs at live action boundaries. The object arm lifts25cm and uses a raised transfer with a reachable wrist tilt before descending to release; transfer requires the complete drawer program and at least14cm opening stable for75 physical ticks. Mixed follows the predeclared complementary parent assignment in meta/shared-scenes100.json and preserves original parent media, state, action and masks. It inherits its parents' independent replay evidence.

observation.arm_active_mask is float32[2] in left/right order. Observation t labels the outgoing 15-physics-tick action interval; task-program and gripper progression is active. observation.arm_motion_mask records coordinate changes over the same interval. Final masks are [0,0]. Both masks were reconstructed from saved controls and per-step cursors. Generation and complete independent replay pass task completion and full-object containment with0.25mm numerical tolerance, state error limits and termination checks. Collisions are allowed and recorded at0.1mm reporting tolerance. At most14 idle physical steps align the first successful completion to an observation. All 150 original videos were decoded and timestamp-checked.

meta/lineage.jsonl records source commits, control hashes, scene indices and immutable parent revisions. Train50 is the predeclared paired subset of its own Train100; only indexing/statistics metadata is rebuilt. Its original parent state/action/mask/media values are retained. The100 seeds, object identities, common50 membership, Sequential orders, Mixed parents and normalized offset samples were frozen before outcomes. Actions are generated directly by the pinned R028-derived live expert; no prior trajectories are spliced or retimed. Shared snapshots and frozen reference clocks, plus control/progress evidence are retained under meta when present.

Lifecycle: reported → audit pending. Execution/publication authorization does not constitute user archival approval. This dataset request does not train a policy.

Expert versions are intentionally selected per scene: working old-expert episodes are retained; only failed scenes use a repaired expert. Whole episodes and their original controls are reused. See direct-expert-lock.json and per-episode lineage.