Deep Reinforcement Learning for Communication-Free Distributed Control of Autonomous Vehicles in Unstructured Intersection.
TL;DR
A novel deep reinforcement learning framework that enables safe and efficient navigation in such communication-free, signal-free, and lane-free intersection environments while meeting stringent safety requirements for practical deployment is proposed.
Abstract
Autonomous intersection without reliance on lane markings, traffic signals, or intervehicle communication remains a critical challenge for decentralized autonomous vehicles (DAVs). In this study, we propose a novel deep reinforcement learning (DRL) framework that enables safe and efficient navigation in such communication-free, signal-free, and lane-free intersection environments. Built upon a continuous proximal policy optimization (CPPO) foundation, our method, CPPO-RA-CL, integrates curriculum learning and a reference-action-guided loss to control AVs' acceleration and steering angle under complex multi-AV scenarios. To enhance perception, we design a multimodal data fusion architecture that combines visual and sensor-based inputs in an ego-centric coordinate system (ECCS). Furthermore, we introduce a rule-embedded hybrid policy to ensure long-term deployment safety by combining learned and fallback control. Extensive simulations demonstrate that our approach achieves robust navigation across diverse geometries, including four-way and three-way unstructured intersections. In particular, long-horizon evaluations over 100million km of cumulative vehicle travel report only 36collisions, corresponding to safety levels consistent with real-world statistics. This work highlights the feasibility of scalable, decentralized AV control using DRL without external coordination or infrastructure support while meeting stringent safety requirements for practical deployment.