Safe exploration of nonlinear dynamical systems: A predictive safety filter for reinforcement learning

Abstract

Despite fast progress in Reinforcement Learning (RL), the transfer into real-world applications is challenged by safety requirements in the presence of physical limitations. This is often due to the fact, that most RL methods do not support explicit consideration of state and input constraints. In this paper, we address this problem for nonlinear systems by introducing a predictive safety filter, which turns a constrained dynamical system into an unconstrained safe system, to which any RL algorithm can be applied `out-of-the-box'. The predictive safety filter receives the proposed learning input and decides, based on the current system state, if it can be safely applied to the real system, or if it has to be modified otherwise. Safety is thereby established by a continuously updated safety policy, which is computed according to a data-driven system model, supporting state and input dependent uncertainties in the prediction.