A Finite-Time Safety Filter for Learning-Based Autonomous Driving

Learning-based control algorithms optimize the driving performance of autonomous vehicles by maximizing cumulative rewards or imitating expert maneuvers. However, these controllers often exhibit worryingly safe performance in the face of real-world applications characterized by constant flux and unpredictability. This paper introduces a finite-time safety filter, for learning-based driving controllers, which can keep the vehicle safe in the presence of unexpected physical constraints. Firstly, we establish a 6-Degree of Freedom (6-DOF) vehicle dynamics model as a predictive tool to calculate the vehicle's target state. An integrated intervention mechanism is designed to adjust the vehicle's target state, ensuring safe distancing from unexpected physical constraints. To minimize disruption to original drivability, we confine the intervention on the vehicle's target state to a constrained time period. Further, we develop a controller capable of generating control inputs based on the post-intervention target vehicle state. Finally, experiments are conducted in the Carsim environment to demonstrate the effectiveness of our safety filter for a reinforcement learning controller. The results indicate that our safety filter can guarantee integrated safety in unpredictable environments with minimal performance penalty.