Learn the Task, Not the Schedule: Reducing Spurious Cross-Arm Coupling in Bimanual Imitation Learning

Key points

• A bimanual demonstration specifies both what the robot should do and one particular schedule for when each arm should act.1
• Standard imitation learning does not distinguish task-required cross-arm constraints from incidental scheduling choices, and may therefore exploit opposite-arm progress as a shortcut cue, creating spurious cross-arm coupling even where relative timing is irrelevant for the current task phase.1
• We use CTR to train Diffusion Policy, Action Chunking with Transformers, and $π{0.5}$, and evaluate them on four simulated and four real-robot tasks.1
• With $π{0.5}$ on an Independent task, a 4 s pause transfers 2.56 s of delay to the opposite arm under the Concurrent schedule distribution, but only 0.25 s with CTR.1

Sources (1)

• [1]Learn the Task, Not the Schedule: Reducing Spurious Cross-Arm Coupling in Bimanual Imitation Learning
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 19, 02:44 PM

  We introduce Constraint-preserving Temporal Recombination (CTR), a training-time data intervention that preserves recorded arm-local sensorimotor content while retiming the corresponding segments to realize admissible schedules that preserve task-required cross-arm constraints.

  A bimanual demonstration specifies both what the robot should do and one particular schedule for when each arm should act.

Extractive summary: sentences quoted from the sources.