A Predictive Safety Filter for Learning-Based Racing Control

The growing need for high-performance controllers in safety-critical applications like autonomous driving motivated the development of formal safety verification techniques. In this letter, we design and implement a predictive filter that is able to maintain vehicle with respect track boundaries when paired alongside any potentially unsafe control signal, such as those found learning-based methods. A model (MPC) framework used create minimally invasive algorithm certifies whether desired input safe can be applied vehicle, or provides an alternate keep bounds. To end, provide principled procedure compute invariant set nonlinear dynamic bicycle models using efficient convex approximation fully support aggressive racing performance without conservative interventions, extended real-time through backup trajectories. Applications assisted manual deep imitation learning on miniature remote-controlled demonstrate filter's ability ensure during maneuvers.