PutCab Unconstrained V4 Train50
Ablation of Shiki42/PutCab-CTR-Train50-V4
at revision 575634eaaaed193458001b41050d0f9f55ecf545: the same 50 scenes and the same recorded per-arm commands,
with the cross-arm precondition removed.
Construction. In CTR, the right (object) arm waits after grasp-and-lift until the drawer is at least 14 cm open. Here that wait is deleted: the right arm's transfer/release program starts immediately after the lift. Every other arm command is taken verbatim from the published CTR episode (no re-planning). Each episode is re-simulated in RoboTwin from its original initial scene in the restored generation environment (bit-identical physics/state/action on an unchanged control episode) and all three cameras are re-rendered.
Outcome (all episodes kept, including failures).
| episodes | task success | |
|---|---|---|
| schedule changed (wait removed) | 26 | 12 |
| unchanged (no wait needed; identical state/action/masks to CTR) | 24 | 24 |
| total | 50 | 36 (72%) |
Typical failures: the right arm reaches the still-closed drawer, collides with its front and blocks the left arm
(6 episodes never open the drawer), or the object is left on the drawer face / cabinet top / table.
Per-episode outcomes, deleted wait, and contact audits are in meta/outcome-summary.json, meta/replay_results/
and meta/schedule_builds/.
Format. LeRobot v2.1, three 320×240 H.264 cameras at 50/3 FPS, 13398 frames.
observation.arm_active_mask follows the same prefix-only IdleMask rule as the parent CTR dataset;
observation.arm_motion_mask records coordinate changes. meta/lineage.jsonl links each episode to its parent.
This is a rebuttal ablation dataset (2026-11-12); it is not a qualified CTR demonstration set.
- Downloads last month
- 95