제어장벽함수와 심층강화학습을 결합한 안전강화학습 제어기 연구

Abstract

In the autonomous driving system, a model-based control approach has issues on the performance degradation and safety due to uncertainties or modeling errors in the used model. Recently, reinforcement learning has received lots of attention as an alternative and attractive technology in autonomous driving, but it also has challenges of how to guarantee safety. Both stability and safety are the most significant components that an autonomous driving system needs for protecting drivers and pedestrians from unexpected accidents. To resolve the above issues, we propose an attractive and efficient control technique by combining a model-free deep reinforcement learning-based controller with a model-based control barrier function to enhance the safety of an autonomous driving system. The proposed control approach builds a safe deep reinforcement learning-based controller by integrating a dynamic model-based control barrier function which guarantees a safe boundary of the vehicle motion. Various simulations are carried out to verify the performance of the proposed RL based control barrier control method, and it is shown that the proposed RL-based controller improved safety by guiding the safe boundary between ego and lead vehicles using barrier function in autonomous driving situations.