Separable Policy Learning for Emergency Vehicle Prioritized Traffic Signal Control

ICLR 2026 Conference Submission13159 Authors

Keywords: Traffic Signal Control, Reinforcement Learning, Separable Policy Learning

Abstract: Traffic Signal Control plays a vital role in optimizing urban traffic flow and reducing accidents by regulating signal phases at intersections. While traditional fixed-time control methods are simple and infrastructure-efficient, they fail to adapt to complex and dynamic traffic patterns, particularly during peak periods or in the presence of emergency vehicles. In this paper, we address the emergency-vehicle-aware traffic signal control problem by proposing a decoupled policy fusion framework that separately optimizes control strategies for regular vehicles and emergency vehicles. The two policies are later combined into a global strategy with automatically learned weights, mitigating the negative impact of $Q$-function approximation errors. We further introduce SplitEMV, a novel multi-agent model that enhances inter-agent communication and decision efficiency. Experiments demonstrate that our method significantly improves emergency vehicle response times while preserving efficiency of regular vehicles. The learned emergency vehicle prioritized policy also integrates seamlessly with existing traffic signal control methods in a zero-shot manner, supporting practical deployment.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 13159