This project addresses the challenge of autonomous motion generation in safety-critical applications. In particular, this project aims to bring safety to unstructured dynamic environments with sudden unknown disturbances. Errors in these kinds of environments can cause severe damages to the system, environment, or even human life. For example, a robot delivering packages in diverse urban environments needs to avoid harming humans. We thus aim to develop a novel framework for robust and safe robot motion in the presence of dynamic uncertainties arising from interaction with a non-stationary environment. To achieve this goal, our framework builds on three tightly integrated and novel formulations of reinforcement learning, adaptive control with robust and fast adaptation, and computer vision-based uncertainty prediction.

2022

2024