Safe Exploration in Model-based Reinforcement Learning using Control Barrier Functions

This paper studies the problem of developing an approximate dynamic programming (ADP) framework for learning online the value function of an infinite-horizon optimal problem while obeying safety constraints expressed as control barrier functions (CBFs). Our approach is facilitated by the development of a novel class of CBFs, termed Lyapunov-like CBFs (LCBFs), that retain the beneficial properties of CBFs for developing minimally-invasive safe control policies while also possessing desirable Lyapunov-like qualities such as positive semi-definiteness. We show how these LCBFs can be used to augment a learning-based control policy so as to guarantee safety and then leverage this approach to develop a safe exploration framework in a model-based reinforcement learning setting. We demonstrate that our developed approach can handle more general safety constraints than state-of-the-art safe ADP methods through a variety of numerical examples.