AHSWN Home · Issue Contents · Forthcoming Papers

Safe Reinforcement Learning for Simulated Pedestrian Collision Avoidance of Autonomous Vehicles
Guangyuan Zou, Ying He, F. Richard Yu, Weike Pan, Zhong Ming and Victor C. M. Leung

Pedestrian collision avoidance is one of the most fundamental problems in autonomous driving, which requires the autonomous vehicles to avoid collisions with traffic participants. The key to this problem is how to appropriately trade-off the traffic efficiency of autonomous vehicles and the unsafety that collision of autonomous driving. Reinforcement learning (RL) provides a promising solution to solve this problem by learning to adapt to pedestrian behaviors. However, traditional methods have low traffic efficiency, and standard reinforcement learning cannot be applied directly due to the weaker safety in training and deployment. To address this limitation, we propose three different safe RL approaches: 1) RL with safe reward, 2) RL with constraints, 3) RL with limited exploration. Our results show that the proposed three Safe RL methods make a better trade-off between the driving efficiency and the unsafety caused by unexpected pedestrian behaviors. The proposed different methods are applicable to different environment setting. RL with safe reward need a well-designed reward function and unrestrict collisions in training. RL with constraints also unrestricted collisions in training, but only need simple reward function. And RL with limited exploration can restricted collisions in training and get a optimal safe policy

Keywords: Safe reinforcement learning, pedestrian collision avoidance, autonomous driving