Task Transfer via Collaborative Manipulation for Insertion Assembly

Insertion tasks are a major difficulty for automatizing manufacturing processes. The bench-mark problem in this category, peg-in-hole, is representative of the challenges that arise in uncertain assembly operations. Humans can carry out this task with ease, even in absence of visual information, relying on haptic and tactile sensing to align the peg with the hole. No generic control approach exist that can reproduce this capability in robots. In this work, we propose a system that allows to transfer this skill to a robot through collaborative task executions. The robot learns from these executions by monitoring its pose and sensed force information and encoding it as a multivariate probability distribution. This distribution is then used to realize an active correction strategy allowing the robot to execute the task autonomously. The approach is validated through a comparative study of peg-in-hole insertion using five different adaptation schemes. The results indicate a significant advantage of the learned model compared to blind random search and insertion without adaptation.