Safe Reinforcement Learning via Control Barrier Functions with Model Uncertainty

A novel framework for safe reinforcement learning employs control-theoretic action perturbations to maintain safety in complex systems with uncertain dynamics. Initially, it develops a probabilistic control-affine dynamics model in an offline setting, subsequently creating control barrier functions (CBFs) that factor in model uncertainty to establish conservative safety constraints. These constraints are implemented via an online action correction mechanism, facilitating safe exploration while minimizing performance limitations. This method overcomes the shortcomings of current safe RL techniques that offer safety assurances only in expectation, as well as control-theoretic methods that depend on known dynamics or precise model estimation. Empirical tests validate the effectiveness of this framework.