Kavin60606/bimanual-handover-isr-std

Bimanual handover — ISR-standardized (agilex_piper_bimanual)

Cleaned + ISR-resampled version of PranayTest/bimanual-handover-2026-08-29 (AgileX Piper dual-arm, 20 fps, state/action[14] = 12 joint angles in degrees + 2 grippers normalized 0-1, cameras top / left-arm / right-arm 480x640).

What was done

1. Episode curation — 6 of 56 episodes discarded ([2, 9, 40, 45, 54, 55]):

ep	reason
2	left arm never participates (left gripper flat) — single-arm, no handover
9	18 s, zero gripper motion on both arms — aborted, never grasps
40	left gripper dead, 58% idle frames — single-arm only
45	9.4 s, zero gripper motion — dead clip
54	29 s, 75% idle, grippers start closed, motion only at tail — truncated attempt
55	left gripper ends closed (1.00 -> 0.04) — episode cut before release, incomplete handover

2. ISR — Information-Standardized Trajectory Resampling (Yang et al., IROS 2026, arXiv:2606.22907, derived from the official D-Robotics-AI-Lab/ISR implementation). Instead of keeping every k-th frame in time, frames are kept at ~equal gaps of accumulated information (spatial motion + acceleration). Operator pauses collapse to nothing; motion and deceleration-before-contact are densely sampled; every gripper open/close window is force-kept.

Knobs (paper values, no data-derived tuning):

knob	value	source
d_target	0.05	paper §III-A — single empirical value, uniform across tasks
lambda_dist	1.0	paper default
lambda_acc	0.01	paper Table III, pick-and-place tasks
gripper_threshold	0.05	official repo code default (0-1 gripper channel)
gripper forcing	either-arm union	frames where either gripper changes are always kept

ISR ran on the 12-D joint vector (degrees). Note: the paper operates on end-effector positions; joint-space application is an extension. Timestamps are restamped on a uniform 20 fps grid (standard LeRobot convention) — the standardized sequence is pacing-normalized, not wall-clock.

Result

	value
episodes	50 (of 56 source)
frames in (clean eps)	102,266
frames out	78,379
kept	76.6%
fps (restamped)	20

Per-episode keep ratios vary ~50-92% (content-adaptive: idle-heavy episodes compress hardest).

Known source-data caveat

The source dataset shows a ~2-step (0.1 s) control delay between action and state (state lags action). If you train on this dataset, consider aligning action[t] with state[t+2] when constructing training pairs.

Built with teleop_std_poc (fetch -> ISR -> materialize); build log in build_results.json.