The paper introduces Quality-Controllable Policy (QCP), a robot policy that lets you set execution speed and motion smoothness at inference time.
Method: An iterative hindsight relabeling loop collects online rollouts under target quality labels, adds successful trajectories to the dataset, relabels them by measured quality, and fine-tunes the policy.
Simulation results: On Robomimic and RoboEval, the paper says QCP improves quality and success rates across iterations, and its successful rollouts exceed the quality of the initial demonstrations.
Real-world tests: Controlling speed increased throughput in conveyor operations, and controlling transport stability improved success when carrying wine glasses.
Sign in to comment.
TL;DR of the paper, generated automatically.
The paper introduces Quality-Controllable Policy (QCP), a robot policy that lets you set execution speed and motion smoothness at inference time.
Speed for the conveyor, stability for the drinks.