Motion Planning at Intersections with Event-driven Recurrent Q-Learning

Abstract

Autonomous driving at intersection has great potential on control for smart cities to relieve the energy consumption and transportation congestion. However, it remains challenging to find promising behavior sequence in multi-agent environment with uncertain participation of obstacles. This work develops Event-driven Recurrent Q-Learning (ERQL) to focus on the motion planning task towards intersection scenarios to conclude a sample path with safety and efficiency. We elaborate the definition of events to capture the environment structure and introduce recurrency to process sequence model. Besides, we incorporate collision-avoidance into the event-driven framework and design a mechanism to extract recurrent feature from replay buffer in Q-learning framework. Simulation results show that the developed off-line learning procedure can adapt to on-line decision making towards uncertain agent behaviors.