Ensemble Probabilistic Model Predictive Safety Certification for Learning-Based Control

Despite the proven efficacy of learning-based control schemes, the absence of safety guarantees prevents their deployment in many applications. This paper addresses this problem by introducing a novel safety certification framework that leverages the principles of stochastic model predictive control. Our proposed framework modifies the actions taken by a learning-based controller in a minimally invasive fashion to enable non-conservative decision-making with adjustable safety levels for systems subject to additive unbounded noise. In addition to that, an ensemble of probabilistic neural networks is used to model the unknown aspects of the system dynamics. In situations where the proposed safety certification fails to find solutions, a constraint satisfaction probability maximizing backup algorithm is proposed. Finally, the efficacy of the proposed safety certification framework in reducing the number of violations is shown through simulations of reinforcement learning within a range of diverse environments.