VisionLight: Dynamic Flow–Driven Uncertainty Modeling in 3D Visual Reinforcement Learning for Traffic Signals
Keywords: VisionLight, reinforcement learning, traffic signal control, video-based traffic management, entropy attention
TL;DR: VisionLight is an RL-based traffic signal control system using video input and entropy attention to improve flow prediction, outperform baseline SOTA, and remain robust under different weather conditions.
Abstract: Most RL-based traffic signal control (TSC) methods rely on features such as vehicle coordinates and waiting times, which are available in simulation but not at real intersections. We present VisionLight, an RL framework that operates on real-time video input through two modes: (1) end-to-end processing of raw footage and (2) image-based feature extraction compatible with existing TSC systems. To address flow-change uncertainty, VisionLight introduces an Entropy Attention & Multi-agent Mechanism tuned for turn-based traffic. It achieves an average 56.8% improvement over SOTA baselines across three metrics, matches feature-driven RL models, and generalizes robustly under extreme weather without retraining, making it practical for real-world deployment.
Primary Area: reinforcement learning
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2026/AuthorGuide.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 8239
Loading