Safe, Efficient, and Comfortable Reinforcement-Learning-Based Car-Following for AVs with an Analytic Safety Guarantee and Dynamic Target Speed

Over the last decade, there has been rising interest in automated driving systems and adaptive cruise control (ACC). Controllers based on reinforcement learning (RL) are particularly promising for autonomous driving, being able to optimize a combination of criteria such as efficiency, stability, comfort. However, RL-based controllers typically offer no safety guarantees. In this paper, we propose SECRM (the Safe, Efficient, Comfortable car-following Model) that balances traffic efficiency maximization jerk minimization, subject hard analytic constraint acceleration. The acceleration is derived from criterion follower vehicle must have sufficient headway be avoid crash if leader brakes suddenly. We critique time-to-collision (TTC) threshold (commonly used RL controllers), confirm simulator experiments representative previous TTC-threshold-based autonomous-vehicle controller may (in both training testing). contrast, verify our safe, scenarios with wide range behaviors, regular-driving emergency-braking test scenarios. find compares favorably comfort, speed-following classical (non-learned) (intelligent driver model, Shladover, Gipps) controller.