Probabilistic Model Predictive Safety Certification for Learning-Based Control

Reinforcement learning (RL) methods have demonstrated their efficiency in simulation. However, many of the applications for which RL offers great potential, such as autonomous driving, are also safety critical and require a certified closed-loop behavior order to meet specifications presence physical constraints. This article introduces concept called probabilistic model predictive certification (PMPSC), can be combined with any algorithm provides provable certificates terms state input chance constraints potentially large-scale systems. The certificate is realized through stochastic tube that safely connects current system terminal set states known safe. A novel formulation allows recursively feasible real-time computation tubes, despite possibly unbounded disturbances. design procedure PMPSC relying on Bayesian inference recent advances invariance presented. Using numerical car simulation, method its illustrated by enhancing an certificates.