Safe Active Dynamics Learning and Control: A Sequential Exploration–Exploitation Framework

Safe deployment of autonomous robots in diverse scenarios requires agents that are capable efficiently adapting to new environments while satisfying constraints. In this article, we propose a practical and theoretically justified approach maintain safety the presence dynamics uncertainty. Our leverages Bayesian meta-learning with last-layer adaptation. The expressiveness neural-network features trained offline, paired efficient online adaptation, enables derivation tight confidence sets, which contract around true as model adapts online. We exploit these sets plan trajectories guarantee system. handles problems high uncertainty, where reaching goal safely is potentially initially infeasible, by first exploring gather data reduce before autonomously xmlns:xlink="http://www.w3.org/1999/xlink">exploiting acquired information perform task. Under reasonable assumptions, prove our framework guarantees high-probability satisfaction all constraints at times jointly, i.e., over total task duration. This theoretical analysis also motivates two regularizers models improve adaptation capabilities well performance reducing size sets. extensively demonstrate simulation on hardware.