Interpretable Vehicular Safety Control Based on Hybrid Automata and Deep Reinforcement Learning

Advanced driver assistance system (ADAS) is an emerging technology in autonomous vehicles. It involves discrete function switches with continuous dynamics, thus sudden changes of different control strategies may lead vehicles to lose stability and safety. While existing analysis mainly focuses on individual functions, and ignores the continuity of control variables when switching between various functions. Meanwhile, interpretable control procedures with safety guarantees are often absent due to the complex vehicular systems. Therefore, verifying ADAS functions with safety guarantees and interpretability is essential, especially for adaptive cruise control which is a main function of ADAS. This paper proposes an approach for safety control based on hybrid automata and deep reinforcement learning. First, hybrid automata are built to describe ADAS functions, and then a deep reinforcement learning model is employed to optimize the strategies for adaptive cruise control. Finally, numerical experiments demonstrate that the proposed approach effectively manipulates vehicle maneuvers across four different scenarios, including acceleration, deceleration, emergency braking and vehicle cut-in. Furthermore, this approach shows strong interpretability, high efficiency, and low resource cost, with an inference requiring 1.08 ms and 456.89 MB of memory on the CPU.