ICON Seminar in Robotics: Prof. Wei Pan (TU Delft)

Recent years have seen a great number of contributions to machine learning methods for real-world robotic control from both the control and machine learning communities. In many applications, how to formally certify the safety of a data- and learning-based control policy is paramount. We present a Bayesian method in reinforcement learning by modeling the control Lyapunov function and control barrier function in control theory as priors. The sporadically used Bayesian method shows a great promise in control performance, stability and safety guarantee, and computational cost. We will illustrate these "safe Bayesian reinforcement learning" methods on both simulated robot control tasks and real-world experiments, including spacecraft autonomous rendezvous and proximity operations and objects pushing using a Clearpath Husky.