LAUSR

Skill generalization in force fields is quite challenging and has not been fully investigated yet in the domain of robot learning. In this letter, we present a novel adaptation strategy that allows a robot to generalize the learned skill to deal with new task conditions with different force fields. A force-relevant skill is represented by a set of parametric compliant profiles, which are initialized during the human demonstration. The motion trajectories captured from demonstration are encoded as dynamical movement primitives (DMPs). Subsequently, the parameters representing the compliant profiles are adapted iteratively to handle the new task situations, based on the motion errors between the output of DMPs and the current movements. We validate our approach in several scenarios, including two numerical simulation tasks (circle tracking and point-to-point movement), one physical simulation task (catching-a-ball), and two real-world tasks (moving-a-plank and plug-into-socket).