Quality and selection
Release checks
The current release contains 1,651 independent recordings, 6,004 MP4 files and 2,381,854 image frames. All retained source image messages were decoded and exported, with zero image decode failures. All 1,651 episodes passed independent checks of payload hashes, calibration references, source/video indexing, exact integer timestamp relations, numerical row counts and optional record-time matching after the scope update on 2026-10-26.
These checks establish export consistency. They do not establish successful task completion, accurate SLAM, accurate physical synchronization or high-precision calibration.
Selection scope and pose alerts
The reviewed candidate pool contained 2,415 canonical operation recordings. The current release retains 1,651 and excludes 764: 714 through whole-batch scope/quality decisions and 50 individually flagged poses. Duplicate copies were not independent candidates. The 100 recordings in batch 0920_pick_cup were removed from the release on 2026-10-26 at the owner's request. Cup picking and hanging-cube exclusions are release-scope decisions, not claims that these recordings failed quality checks. Historical full-export reports describe the earlier 1,751-record selection.
The 50 individually excluded records were five from 0805 picking, 34 from 0810 drawer, one from 0814 drawer and ten from 0828 drawer. Existing alerts used a sampled 10 Hz translation step of at least 0.20 m in the relevant audit. The 0825 picking and drawer batches were excluded as whole batches after severe pose failures.
The early 0727/0728 batches were accepted through project review rather than a newly rerun uniform SLAM audit. The later 0923/0925 batches had no identified rejection record during selection, which is not a fresh all-trajectory SLAM accuracy test. Per-episode pose_review_status retains these distinctions.
Task outcome and missing observations
100 button records carry the label button_did_not_light_user_report. The remaining 1,551 records are not individually outcome-labelled. Do not compute a success rate by treating unknown labels as successful. Drawer sequences can contain closing and regrasping; the export preserves the complete captured record rather than a training-specific crop.
1,451 episodes have all four image streams; 200 have main RGB only. All retain recorded SLAM pose and gripper streams. Missing auxiliary videos are intentional, not decode failures. Left and right image counts differ by 17 and should not be paired by equal frame number alone.
Encoding checks
Captured JPEG images were re-encoded with libx264, CRF 18, fast preset, yuv420p, no B-frames and maximum GOP 30. No geometric resize, undistortion or crop was applied. Every output video was decoded during export to verify display-frame count, dimensions and PTS. The scope update did not re-encode retained videos.
For the retained selection, 22,881 deterministic source/export frame comparisons have PSNR from 38.21 to 47.67 dB, median 40.51 dB. Median mean absolute pixel error is 2.052 on an 8-bit scale. Sampling used video-frame indices 0,30,100,300,600,1000 when available. These describe compression differences, not a view-synthesis benchmark or an exhaustive perceptual-quality guarantee.
The maximum encoding-PTS adjustment is 500 ns; original timestamps are retained. No capture synchronization accuracy is claimed. The 1403 main–USB calibration limitation is documented in CALIBRATION.md.
Statistical definitions
Recorded duration is summed bag end-minus-start time, including margins. Per-stream average rates are (message_count−1)/(last−first record timestamp), per episode. They are not hardware FPS measurements. Episode payload size includes files inside episode directories, excluding root metadata, tools, QA images and color-chart attachments. See statistics.json for exact values.