Improving Robotic Imitation Learning via Trajectory Standardization
Paper • 2606.22907 • Published
ISR-standardized bimanual teleop dataset. Source PranayTest/bimanual-handover-2026-08-26
(agilex_piper_bimanual, 2 arms, task right_pick_handover_left_place) run through
ISR — Information-Standardized Trajectory Resampling (Yang et al., IROS 2026,
arXiv:2606.22907) to remove teleop pacing noise
(operator pauses / variable speed) while preserving reach + contact dynamics.
| field | value |
|---|---|
| episodes | 86 |
| fps | 20 |
| total frames (standardized) | 138,120 |
| total duration (standardized) | 6,906.0 s (115.1 min) |
| source frames | 182,448 |
| frames kept | 75.7% |
| cameras | top, left-arm, right-arm (480x640 @ 20fps) |
| state / action | 14-dim (R j1-6, L j1-6, R grip, L grip) |
d_target = 0.5 (info-gap target; joints in degrees, rescaled from paper EE default)lambda_acc = 0.01 (acceleration weight — preserves decel-before-contact)lambda_dist = 1.0, gripper_threshold = 0.05 (force-keep on EITHER arm's grasp)LeRobot v3.0-style, packaged in mrfood_data_standardised.tar.gz:
mrfood_std/
data/chunk-000/file-000.parquet # kept rows: observation.state, action, timestamp, frame_index, episode_index, index, task_index
videos/observation.images.{top,left-arm,right-arm}/chunk-000/file-000.mp4 # re-encoded to kept frames
meta/info.json meta/episodes/... meta/tasks.parquet METADATA.json
Timestamps are re-based uniform (1/fps) — motion-standardized playback; ISR removed the pause frames,
so wall-clock is shorter than the raw recording. See meta/info.json -> isr_standardization.