A predictive safety filter for learning-based control of constrained nonlinear dynamical systems

The transfer of reinforcement learning (RL) techniques into real-world applications is challenged by safety requirements in the presence physical limitations. Most RL methods, particular most popular algorithms, do not support explicit consideration state and input constraints. In this paper, we address problem for nonlinear systems with continuous spaces introducing a predictive filter, which able to turn constrained dynamical system an unconstrained safe any algorithm can be applied ‘out-of-the-box’. filter receives proposed control decides, based on current state, if it safely real system, or has modified otherwise. Safety thereby established continuously updated policy, model formulation using data-driven considering dependent uncertainties.