Power Flow Regulation, Adaptation, and Learning for Intrinsically Robust Virtual Energy Tanks

Ideally, a robot controller should not only be designed to exhibit a given interaction behavior under controlled conditions, but also to be robust to changes e.g. in the environment. Within the paradigm of virtual energy tanks for passivity-based controls, robustness may be provided by setting absolute limits on the tank energy. However, an energy limit alone does not prevent a sudden drain of the tank, which may result in a sudden increase of potentially problematic, passivity-violating energy somewhere in the system. In this letter, we tackle this problem by regulating the exchanged power between the energy tank and the system according to a reference power trajectory. We propose a method to encode this trajectory and to conservatively learn the corresponding parameters. The resulting system is adaptable and robust to both predicted and unpredicted changes, either in the environment or the system. Experimental results with a Franka Emika Panda robot performing an exemplary force-based interaction task validate the performance improvement with our method.