Force-Based Learning of Variable Impedance Skills for Robotic Manipulation

Numerous robotics tasks involve complex physical interactions with the environment, where the role of variable impedance skills and the information of contact forces are crucial for successful performance. The dynamicity of our environments demands robots to adapt their manipulation skills to a large variety of situations, where learning capabilities are necessary. In this context, we propose a framework to teach a robot to perform manipulation tasks by integrating force sensing and variable impedance control. This framework endows robots with force-based variable stiffness skills that become relevant when vision information is unavailable or uninformative. Such skills are built on stiffness estimations that are computed from human demonstrations, which are then used along with sensed forces, to encode a probabilistic model of the robot skill. The resulting model is later used to retrieve time-varying stiffness profiles. We study two different stiffness representations based on (i) Cholesky decomposition, and (ii) Riemannian manifolds. For validation, we use a simulation of a 2D mass-spring-damper system subject to external forces, and a real experiment where a 7-DoF robot learns to perform a valve-turning task by varying its Cartesian stiffness.