Event

Learning a Speed Prior for Robot Demonstration Retiming

1 reports · 1 sources

Seungyong Moon and co-authors propose learning a speed prior from large-scale demonstration data to retime robot demonstrations. Given a task instruction, a motion path, and images along the way, the model predicts a distribution of execution speeds along the path and uses a selected quantile as a unified clock to retime each demonstration. The authors state that this speed prior is fit on robot teleoperation data and on handheld demonstrations in which a human hand, rather than a robotic arm, moves the gripper. At execution time, the policy works with a real-time optimal control layer to track predicted action chunks at the predicted tempo under robot constraints. The authors report that on 120 RoboLab tasks, policies fine-tuned with the learned clock achieve higher success rates than policies using the original demonstration timing uniformly accelerated to the same speed, with jitter reduced by up to half; on a real robot, cup stacking takes 3.50 seconds, 1.72 times faster than the default policy's 6.02 seconds, with comparable success rates, succeeding in 29 of 30 trials.