ADAS-RL: Safety Learning Approach for Stable Autonomous Driving

Dongsu Lee, Minhae Kwon

Abstract: Stability is the most significant component of an autonomous driving system, affecting both the lives of drivers and pedestrians and traffic flow. Reinforcement learning (RL) is a representative technology used in autonomous driving, but it has challenges because it is based on trial and error. In this letter, we propose an efficient learning approach for stable autonomous driving. The proposed deep reinforcement learning based approach can address the partially observable scenario in mixed traffic which includes both autonomous vehicles and human-driven vehicles. Simulation results show that the proposed model outperforms the control-theoretic and vanilla RL approaches. Furthermore, we confirm the effect of the sync-penalty, which teaches the agent about unsafe decisions without experiencing the accidents.