Event

QCP: Controllable Robot Policy Quality at Inference Time

1 reports · 1 sources

Sinjae Kang and other authors propose Quality-Controllable Policy (QCP), which can control individual quality attributes such as execution speed and motion smoothness, as well as their combinations, at inference time. To extend the controllable quality range beyond the initial demonstrations, the authors designed an iterative post-hoc relabeling loop: collect online rollouts according to target quality labels, add successful trajectories to the dataset, then post-hoc relabel the accumulated trajectories according to measured quality and fine-tune the policy. The authors report that experiments on Robomimic and RoboEval show that QCP continuously improves task execution quality and success rate over iterations, and that the quality of the successful trajectories it produces exceeds that of the initial demonstrations; in real manipulation tasks, controlling speed increased throughput in conveyor-belt operations, and controlling transport stability improved the success rate when carrying a wine glass.