Integrated decision and control at multi-lane intersections with mixed traffic flow

Autonomous driving at intersections is one of the most complicated and accident-prone traffic scenarios, especially with mixed participants such as vehicles, bicycles pedestrians. The policy should make safe decisions to handle dynamic conditions meet requirements on-board computation. However, current researches focuses on simplified considering only surrounding vehicles idealized lights. This paper improves integrated decision control framework develops a learning-based algorithm deal complex flows, which can not take account realistic characteristics lights, but also learn under different safety constraints. We first consider velocity models for green red lights in training process use finite state machine modes light transformation. Then we design types distance constraints pedestrians, respectively formulize constrained optimal problems (OCPs) be optimized. Finally, reinforcement learning (RL) value networks adopted solve series OCPs. In order verify efficiency proposed method, multi-lane intersection existence large-scale set practical phases. simulation results indicate that trained well balance tracking performance. Compared model predictive (MPC), computational time three orders magnitude lower.