Safe reinforcement learning via a model predictive control approach

Schäfer et al. propose a safe reinforcement learning framework that uses MPC as an offline safety oracle to precompute feasible state-action spaces, keeping exploration of robots and cyber-physical systems within mathematically verified bounds. The method maps convex intervals of continuous action spaces via grid search plus binary search, and uses a projection filter to intercept unsafe actions. The authors validated it on a Quanser Aero 2 single-degree-of-freedom pitch platform. Supplementary notes: - The work targets the risk of unconstrained exploration in standard RL on high-inertia hardware, defining irreversible "critical points"—i.e., feasible set boundaries. - Safety determination relies on the MPC optimizer finding a valid action sequence within a future prediction horizon, and the horizon must exceed the system's stopping distance to ensure recursive feasibility.