DCC1015/piper_handout_merged_v4_v5

Piper paper-cup handover — 79 curated episodes

Task: Pick up the paper cup with the gripper, place it in the person’s hand, and then open the gripper.

This LeRobot v2.1 dataset contains 79 episodes, 38,948 data frames and 158 RGB videos at 30 FPS, 640×480. Episode indices are 0–78, with continuous global frame indices 0–38,947. All episodes are in the training split.

Source	Retained episodes	Frames	Current episode indices
piper_handout_v4	21	10,023	0–20
piper_handout_v5	41	20,745	21–61
piper_handout_v6	17	8,180	62–78
Total	79	38,948	0–78

Cleanup performed on 2026-09-15

All original episode IDs in this section refer to the previous 85-episode merged dataset, using zero-based numbering.

• Removed original episodes 0, 1, 2, 63, 77 and 78 as requested.
• Trimmed the first 129 frames (4.30 seconds) of original episode 24, now episode 21. Both camera streams and all data columns use source frames [129,733).
• Removed unreferenced video tails listed below. Their data values and lengths are retained.

Original episode	Current episode	Kept frames per camera	Removed tail frames per camera
3	0	553	608
48	45	435	18
62	59	684	212
66	62	395	199
83	77	459	100

Every video now has exactly the corresponding number of data rows. Retained action/state values are unchanged. During that cleanup, untrimmed videos were copied byte-for-byte and retained decoded pixels of trimmed videos were checked, before the later color correction below. Timestamps, indices, per-episode statistics and aggregate statistics were updated.

The original v6 static/incomplete filter is recorded for historical reference in meta/v6_stationarity_report.json; its IDs refer to the original v6 recording, not current merged indices.

See meta/cleanup_manifest.json for the complete old-to-new mapping and trim bounds, meta/merge_manifest.json for source provenance, and meta/merge_validation.json for current validation. The prior full local dataset is preserved in /home/a/piper_data/piper_handout_merged_v4_v5_v6.before_cleanup_20260915.

This curated, RGB-corrected dataset is published in DCC1015/piper_handout_merged_v4_v5, including the retained v6 episodes. The repository name is kept for compatibility.

RGB correction on 2026-09-15

All 158 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/piper_data/piper_handout_merged_v4_v5_v6.before_rgb_fix_20260915. See meta/color_correction.json for the transformation and per-video verification, and meta/merge_validation.json for current validation.