Traffic-VL: Reinforced Vision-Language Models for Generalizable End-to-End Traffic Signal Control

Traffic-VL introduces a vision‑language model that directly maps synchronized multi‑view images and natural‑language phase descriptions to traffic signal decisions, bypassing intermediate structured states. It is trained using SynTra‑V, a photorealistic simulation that generates surveillance imagery from physics‑based traffic models, and employs offline reinforcement learning to adapt its vision encoder before freezing it for online control. Experiments demonstrate that Traffic‑VL matches the performance of structured‑state controllers while outperforming image‑based and general‑purpose vision‑language models, and it generalizes robustly to unseen topologies, degraded sensing, real footage, and real‑world deployment across multiple intersections.

Machine-generated by The Flow from the publisher's headline and feed description — not written or checked by a human. The full article lives at arXiv AI.

• llms
• reinforcement-learning
• multimodal