Shiki42/ctr-scan-object-original-20260919

Scan Object — Original native expert, fixed training scenes

50 accepted trajectories, 11238 state/action samples, from the existing50 training scenes. The initial attempt accepted47/50; user-authorized planner-only retries recovered the remaining3 without replacing any scene. This is a newly simulated Original dataset using official RoboTwin scan_object.play_once and unchanged native motion controls. It is not a renamed Concurrent, timed, CTR or Mixed dataset. IdleMask is disabled; no idle-mask features are included.

Coverage and failures

Attempted:50. Accepted:50. Failed:0. Failed scene seeds: None. No replacement seeds were used. Planner-only retry attempts: 5; total planning attempts including the initial50: 55. Every retry holds the scene seed, object identity, initial snapshot and native expert geometry fixed; only planner random seeds vary, stopping at first success. See meta/collection_summary.json for exact ordered training seeds and failures. The50-scene training cohort is disjoint from E257 Scan Native100 and its complete 300-candidate pool (400001–400300), as selected by the user. All50 fixed training scenes are represented by one successful trajectory each. Initial failures and all retry outcomes remain recorded.

Capture and schema

Aloha AgileX,250Hz physics,25FPS, synchronized320×240 head and dual wrist RGB. Both wrists use centered_fovy90,90° vertical FOV and the calibrated centered optical-axis mount, RT32 rendering. Actual per-frame intrinsics and mount transforms were validated. Original scene actor configurations, poses and robot qpos match the existing parent snapshot with absolute tolerance1e-7 before planning. Native grasp/lift/object-position/scan stages are unchanged.

observation.state records measured joint angles and measured normalized gripper positions. action is the control target at the next sample: state[t] → target[t+1]. The final observation supplies a target/validation frame and is not an extra training row. Three video streams decode completely and were checked against per-episode frame timestamps. LeRobot v3 numeric rows match raw HDF5 exactly at float32 precision. Source seed and slot are stored for each row.

Numeric quantiles in meta/stats.json use exact global per-feature samples with NumPy linear interpolation, not averages of episode quantiles. See meta/global_quantile_receipt.json for sample identities, constant dimensions and out-of-range fractions. These are dataset metadata, not trained-model normalization assets. No training or policy evaluation was performed.

Provenance

Experiment E351; R001 recorded the first16 attempts, and R002 continued only the remaining34 unattempted scenes using two processes. Initial failures were75/76/151. R003 then retried only those three under explicit user authorization, preserving all47 existing successful trajectories. All raw data and Run receipts are preserved on the collection host; the published tree contains the qualified LeRobot dataset.

Official upstream RoboTwin: c3ddfa8b97d5519efa828b075999bd0006778e5e. Native task source SHA256: 860dfe2cc86d180096089bc87502c7904f7ce2efb64fb32620c1720ca0178c2b. Initial capture source commit: ded2cfb6ccf2152e17dbc393df3dacd7f6d351e4. Retry capture source commit:9d792fa15436d001790a70f44fab757aeb6f2fcd. Actual left/right IK and graph planner seed binding was checked for retries. Input snapshot metadata SHA256: a18f9069087c62d4db657f6a80afdf5d8a85ac24088b9db80af04c3e994ebe47. The revision marker and SHA256SUMS identify the delivered files.

Lifecycle: reported → audit pending. Automated qualification passed; this does not claim user-approved archival/confirmation or learned-policy success.