In robotics, the energy-aware framework has proven to be very effective for guaranteeing low-level system theoretical properties like passivity and stability. However, once a complex task that includes motion in an unstructured environment with possible interaction has to be performed, energy budgeting strategies have to be implemented at a decision level that might break down the aforementioned properties. In particular safety issues are addressed on a case-by-case analysis of the task with dedicated protocols.
This thesis explores the possibilities arising from embedding reinforcement learning loops that are able to manage both the performance maximization along a complex task while preserving low-level safety and passivity requirements of a controlled robot. The AI module that will be implemented in the control will make the system more flexible with respect to unforeseen events, without the need of considering worst-case scenarios in an off-line planning phase. At the same time, the energy-based framework will be preserved, and physical-based system theoretical properties will keep importance.