sy_01

82 recorded Piper barcode checkout demonstrations in LeRobot v2.1 format: the original 32 episodes from piper_checkout_v3, followed by 50 episodes from piper_checkout_v9.

Task: Pick up a milk carton or a snack from the box, turn the item so that its barcode side faces the separate checkout camera, hold it briefly with the barcode fully visible for scanning, and then put it down in the designated area.

Property	Value
Episodes	82, indexed 0–81
Frames	82,121
Recording rate	30 FPS
Robot type	piper_follower
Actions and robot state	Six joint positions and gripper position
Camera observations	observation.images.front, observation.images.global
Videos	164 H.264 MP4 files, 640 × 480
Split	All episodes in train
Format	LeRobot v2.1

data/ contains one Parquet file per episode; videos/ contains the two camera recordings; meta/ contains tasks, lengths, statistics, and source episode mappings.

Retained action/state values are preserved. Leading waits in episodes 51 and 56 were shortened on 2026-09-15, retaining 0.5 seconds before the first observed motion. Both camera videos were trimmed to the same exact frame boundaries; retained frames were preserved during that trim, before the later color correction below. Timestamps, frame indices, continuous global indices, and affected statistics were updated. Gripper values remain continuous; no binarization is applied.

Excluded source episode indices in piper_checkout_v9: [11]. The validation findings and complete source mapping are recorded in meta/merge_provenance.json. Source recordings remain unchanged locally.

Local trim details: meta/trim_provenance.json. Episode 51: remove 120 frames (4.00 seconds); episode 56: remove 157 frames (5.23 seconds). Total: 82,121 frames across 82 episodes.

RGB correction on 2026-09-15

All 164 camera videos had reversed red/blue channels. Every decoded frame has been corrected by swapping R and B. Frame counts, timestamps, episode numbering, previous trims, and all Parquet data remain unchanged. Per-episode image statistics and aggregate statistics were recalculated from the corrected videos. Files remain H.264 YUV420P at 30 FPS and 640×480. The RGB-to-YUV conversion uses chroma subsampling; H.264 is encoded with CRF 0 and the decoded YUV pixels exactly match the corrected frames provided to the encoder.

The complete pre-correction dataset is backed up at /home/a/hf_merge_sy_01_20260914/sy_01.before_rgb_fix_20260915. See meta/color_correction.json for the transformation and per-video verification, and meta/merge_validation.json for current validation.