Online Learning of Motion Constraints from Human Corrections during Ergodic Surface Finishing

Kurzfassung

Surface finishing - such as sanding, polishing, and grinding - is a time-consuming and health-critical process. Robotic automation offers a promising solution to this challenge, enabling consistent surface coverage and tool control. Ergodic control addresses this contact-rich task by covering target distributions in an online manner, accounting for coverage history. We extend our prior ergodic control framework for surface finishing to enable online learning from human interactions during task execution. A human operator can physically intervene via two complementary modes: corrective guidance applied during contact update motion constraints, while repositioning actions update the desired coverage distribution by indicating where more or less attention is needed. This closes the loop between human feedback and robot learning, transforming the system from offline learning to fully interactive human-robot collaboration, where the robot continuously refines both where and how it works. A video demonstration showing online correction of motion preferences on a real workpiece is available at https://youtu.be/kXhaz11jTac.