piperx-sortletter-0917-70ep-raw

70 MINT letter sorting episodes recorded with EvoMind on September 17, 2026 (UTC+8), with leading both-arm-static frames trimmed.

LeRobot v3.0; bi_piperx_follower; 30 FPS; 88,725 frames; 2,957.500 seconds. Three original 640x480 RGB camera views: left_wrist, right_wrist, right_environment_1. Action and observation.state each contain 14 values (six joints and one gripper per arm).

Cleaning history

Five recording batches named sort_letter_260917, starting at 22:05:49, 22:38:05, 22:59:35, 23:05:11 and 23:21:24, were merged in chronological order. Their episode counts are 20, 20, 5, 17 and 8. Episodes were renumbered 0–69. The earlier 22:01:44 accidental recording was deleted at the user's request before this export and is not included. No episodes from these five retained batches were removed.

The same Trim workflow documented by the reference dataset was applied, removing 963 leading and 0 terminal frames from 89,688 original frames. Static detection uses action differences at 30 FPS, with joint speed below 0.1 degrees/second and gripper speed below 0.1 mm/second. The default terminal retention is 2 seconds. No internal segments were deleted and neither arm was retimed.

All three cameras and robot data use the same contiguous per-episode slice. Action, state and task values are preserved exactly within those slices. Frame/global indices and timestamps are reset. Videos are encoded by the Trim workflow as H.264/yuv420p, CRF 20, veryfast, at their original resolution and FPS. All 210 output videos have their frame counts and timestamps validated; five sampled frames per video are compared to the corresponding source pixels, allowing for lossy re-encoding. All leading static-frame counts are zero when the Trim analysis is rerun.

See trim_manifest.json, provenance.json and VALIDATION_RECEIPT.json for intervals, original-to-current episode mapping, numerical and video verification, and file hashes. Original task text is preserved. This verification checks data integrity and synchronization; it does not label task success or certify robot motion safety.