BEVLight: Generalizable Traffic Signal Control from UAV Bird's-Eye Views

Maonan Wang; Yuxin Cai; Junming Liu; Zhengyan Huang; Yi Yu; Ding Wang; Yifeng Zhang; Yirong Chen

eess.SYOct 29, 2026

BEVLight: Generalizable Traffic Signal Control from UAV Bird's-Eye Views

Authors: Maonan Wang, Yuxin Cai, Junming Liu, Zhengyan Huang, Yi Yu, Ding Wang, Yifeng Zhang, Yirong Chen

Organizations: The Chinese University of Hong Kong, Shenzhen · Shanghai AI Lab · Nanyang Technological University · National University of Singapore

Abstract

Learning-based traffic signal control (TSC) has achieved promising performance, but most methods rely on structured traffic states that must be estimated from sensor observations in real deployments. Visual TSC learns signal decisions directly from images, yet existing methods depend on fixed roadside cameras and remain tied to specific intersection configurations and visual domains. UAVs offer on-demand bird's-eye-view (BEV) observations that jointly capture intersection topology and traffic dynamics, making them a promising sensing platform for visual TSC. We present BEVLight, a generalizable TSC framework that directly maps BEV observations to signal actions. BEVLight grounds visual features onto individual lanes and composes them along the junction topology into movement and phase representations, allowing a shared policy to score candidate phases across heterogeneous intersection layouts and signal plans. To train the policy, we develop a web-based virtual teleoperation platform and collect over 5K human control decisions together with their underlying traffic states. Simulator-provided traffic states first supervise lane- and movement-level representations, after which human phase selections and rankings supervise the phase scorer. Re-rendering the recorded trajectories under diverse appearances preserves both forms of supervision and encourages robustness to visual changes. Closed-loop experiments demonstrate zero-shot generalization across traffic demands, signal plans, and intersections: BEVLight reduces average waiting time by 7--9% relative to perception-to-control visual baselines and matches or outperforms structured-state RL controllers that receive ground-truth traffic states, achieving the lowest waiting time in 12 of 16 scenarios. An offline study further shows that BEVLight transfers to real UAV observations without target-domain adaptation.

Figures & tables

Figure 1: (a) UAV-based visual TSC: BEVLight maps BEV observations to signal phases and is trained from human demonstrations that select and rank feasible phases. (b) The three generalization dimensions considered in this work: structure, demand, and domain.

Figure 2: Illustration of a four-way intersection under two phase configurations, each comprising four traffic signal phases.

Figure 3: Overview of BEVLight. Visual patch features are grounded onto individual lanes using lane masks with content-adaptive pooling (BEV-to-Lane Grounding). Lane representations are composed along the junction topology into movement and phase representations, and a shared scorer ranks candidate phases (Topology-Grounded Composition & Score). Simulator-provided traffic states supervise lane and movement representations, while human phase selections and rankings supervise the scorer; recorded trajectories are re-rendered under diverse appearances with unchanged supervision (Human Demonstration for Training).

Figure 4: Eight junctions, each with two signal plans. Blue borders mark plans used for training, while purple borders mark held-out plans.

Unseen Demand
	Conventional		Structured-State RL				Vision-Based
Junction / Plan	FixedTime	MaxPressure	UniTSA	IntelliLight	AttendLight	MPLight	YOLO-C	BEV-S	BEVLight
Beishahe A	18.64 $\pm\,0.39$	8.02 $\pm\,1.87$	11.67 $\pm\,0.62$	7.45 † $\pm\,0.06$	8.08 $\pm\,0.68$	8.26 $\pm\,1.24$	8.01 $\pm\,2.07$	7.73 ‡ $\pm\,0.68$	7.40 ∗ $\pm\,0.23$
Beishahe B	18.30 $\pm\,1.24$	8.01 $\pm\,1.81$	10.21 $\pm\,1.38$	6.68 ‡ $\pm\,1.32$	6.83 $\pm\,1.22$	6.67 † $\pm\,0.93$	7.91 $\pm\,1.76$	7.57 $\pm\,2.07$	6.62 ∗ $\pm\,0.60$
Gaojiaoyuan A	26.09 $\pm\,3.99$	10.20 $\pm\,2.78$	13.04 $\pm\,4.15$	10.55 $\pm\,3.25$	10.28 $\pm\,4.90$	10.71 $\pm\,4.38$	9.55 † $\pm\,4.06$	10.08 ‡ $\pm\,4.45$	9.12 ∗ $\pm\,2.31$
Gaojiaoyuan B	25.12 $\pm\,3.01$	10.37 $\pm\,1.38$	11.69 $\pm\,4.13$	9.34 $\pm\,0.43$	9.18 ∗ $\pm\,2.06$	10.18 $\pm\,1.81$	9.24 † $\pm\,0.25$	10.07 $\pm\,0.85$	9.24 † $\pm\,1.21$

Table 1: Closed-loop average vehicle waiting time (s) across three generalization settings. Values are mean

\pm

sample standard deviation over the two held-out demand patterns. Lower is better. ∗ , † , and ‡ denote the first-, second-, and third-best results in each row, respectively.

Components			Guanghua		Songdo
Adaptive Pooling	Visual Pretraining	Soft Loss	A	B	A	B
$\times$	$\times$	$\times$	9.20	10.73	13.96	23.79
$\checkmark$	$\times$	$\times$	9.38	10.91	13.08	29.03
$\checkmark$	$\checkmark$	$\times$	8.78	9.12	12.86	18.20
$\checkmark$	$\checkmark$	$\checkmark$	7.95	9.04	12.77	18.10

Table 2: Ablation results in average waiting time (s).

Figure 5: Phase predictions on real UAV imagery from Songdo intersections.

Method	Params (M)	GFLOPs	Latency (ms) $\downarrow$
YOLO11n	2.662	26.74	18.44
BEV-S	1.260	37.61	8.60
BEVLight	1.260	37.61	8.59

Table 3: Model size and full inference latency.

Appendix figures & tables13 assets

Supplementary material from the paper’s appendix.

Appendix

Figure 6: BEV observations and lane masks for the eight simulated intersections. For each scene, the lane masks indicate the visible lane regions used by BEVLight, with different colors denoting individual lanes.

	Ground coverage		Visible lane length (m)
Intersection	Footprint (m $\times$ m)	Area (m 2 )	Min.	Mean	Max.
Gaojiaoyuan	$140.86\times 140.86$	19,842	58.13	60.00	62.09
Beishahe	$131.48\times 131.48$	17,287	50.82	57.90	62.87
Pinganli	$147.89\times 147.89$	21,871	55.22	58.27	60.67
Yongrunlu	$147.50\times 147.50$	21,756	56.59	59.09	61.20
Zhijingdao	$111.95\times 111.95$	12,533	56.01	60.13	64.10

Appendix

Table 4: Spatial coverage of the eight simulated intersections.

Figure 7: Traffic-demand profiles for the eight simulated scenes.

	Training			Held-out Evaluation
Scene	Off-peak	Commuter fluctuation	Peak-hour	Demand build-up	Irregular traffic
Gaojiaoyuan	$20.6\pm 1.5$	$33.0\pm 3.0$	$48.0\pm 1.4$	$39.8\pm 9.8$	$34.4\pm 3.3$
Beishahe	$21.0\pm 1.4$	$34.0\pm 6.0$	$50.4\pm 1.9$	$44.0\pm 11.2$	$42.0\pm 7.4$
Pinganli	$26.0\pm 2.1$	$38.8\pm 3.7$	$54.8\pm 3.0$	$43.6\pm 10.3$	$43.6\pm 3.9$
Yongrunlu	$22.6\pm 1.0$	$39.8\pm 3.3$	$56.0\pm 1.3$	$39.8\pm 9.8$	$37.6\pm 3.9$
Zhijingdao	$20.8\pm 1.6$	$31.4\pm 2.5$	$45.6\pm 1.5$	$33.2\pm 7.7$	$36.0\pm 3.6$

Appendix

Table 5: Traffic-demand statistics across the eight simulated scenes. Each entry reports the mean

\pm

standard deviation of the vehicle inflow over five time intervals, in vehicles per minute.

Figure 8: Real UAV observations and corresponding lane masks for intersections O, P, T, and H.

	Ground coverage		Visible lane length (m)
Intersection	Footprint (m $\times$ m)	Area (m 2 )	Min.	Mean	Max.
Songdo, O	$118.5\times 118.5$	14,050	25.61	30.85	37.90
Songdo, P	$119.3\times 119.3$	14,239	22.28	27.11	32.90
Songdo, T	$120.7\times 120.7$	14,568	14.53	25.53	43.24
Songdo, H	$121.2\times 121.2$	14,695	17.54	25.06	51.38

Appendix

Table 6: Spatial coverage of the four real-UAV intersections.

Figure 9: Web-based virtual TSC platform for collecting human demonstrations. The interface displays the current bird’s-eye-view traffic scene and all feasible signal actions. Hovering over an action highlights its corresponding lanes and permitted movements on the intersection.

Panel of Figure 10 — Figure 10: Appearance-diverse rendering of recorded human demonstration states. For each intersection, the top-left panel visualizes the recorded vehicle configuration, while the remaining panels show the same traffic state under five visual conditions: Morning , Noon , Dusk , Dusk-Rain , and Afternoon-Rain . The underlying traffic state and human action ranking remain unchanged across all rendered variants. The examples shown here correspond to Beishahe, Beijing and Gaojiaoyuan, Beijing .

Panel of Figure 11 — Figure 11: Appearance-diverse rendering examples for Pinganli, Beijing and Yongrunlu, Beijing .

Panel of Figure 12 — Figure 12: Appearance-diverse rendering examples for Guanghua, Chengdu and Yau Ma Tei, Hongkong .

Panel of Figure 13 — Figure 13: Appearance-diverse rendering examples for Songdo, SouthKorea and zhijingdao, Tianjin .

Panel of Figure 14 — Figure 14: Representative examples of human control preferences and action rankings. Each example shows a traffic observation together with the ranking assigned by the human operator over all feasible signal phases. The ranking preserves the relative ordering among feasible actions rather than retaining only the top-ranked phase, providing richer supervision for learning the phase-scoring function in BEVLight.

Figure 15: Zero-shot phase predictions over six consecutive UAV frames at Junctions O, P, T, and H in Songdo. Predictions are evaluated offline without controlling actual signals.

Traffic Signal Control

Traffic signal control studies how to set intersection phases and green times to reduce congestion while balancing safety, emissions, and priority needs.

See all 41 papers

Sep 29, 2026

VLALight: A Vision-Language-Action Model for Traffic Signal Control

Traffic signal control (TSC) is essential for improving urban mobility and reducing congestion. Although roadside cameras are widely deployed at signalized intersections and provide rich visual observations of evolving traffic, existing TSC methods typically rely on manually engineered traffic states or separate perception modules, creating a gap between physical observations and control decisions. We present VLALight, the first vision-language-action (VLA) model for end-to-end traffic signal control from multi-view roadside videos. VLALight directly maps visual observations to coordinated signal actions through multi-target spatiotemporal traffic reasoning and topology-aware cooperative perception across intersections. To establish this capability, we develop a two-stage supervised cold-start training strategy for visual traffic understanding and signal decision-making, followed by cooperative agentic reinforcement learning that jointly optimizes local control and network-wide traffic efficiency. Furthermore, VLALight introduces adaptive fast and slow reasoning modes, enabling the policy to allocate deeper reasoning only when additional deliberation provides sufficient control benefits. Through balanced mode-aware rollouts and relative advantage optimization, VLALight learns to trade off decision quality and inference cost. Extensive experiments on seven real-world traffic-flow datasets across three urban networks demonstrate that VLALight consistently outperforms transportation-based, RL-based, and LLM/VLM-based baselines. Ablation studies validate the effectiveness of cooperative perception, network-level optimization, and adaptive reasoning. These results demonstrate the potential of VLA models for real-world physical traffic control. Our project is available at https://github.com/usail-hkust/VLALight.git.

Pan Zhang, Siqi Lai, Kemu Dong +1

The Hong Kong University of Science and Technology (Guangzhou) · Dalian University of Technology

Intelligent Transportation Systems Vision-Language-Action Models Traffic Signal Control

Ask PDF

Sep 25, 2026

VLALight: Lightweight Vision-Language-Action Models for Emergency-Aware Traffic Signal Control

Traffic signal control (TSC) is essential for mitigating urban congestion. Recent advances in vision-language models (VLMs) enable richer interpretation of intersection scenes, opening new opportunities for visual-context-aware TSC. However, the loose coupling and repeated information conversion between modules can lead to the loss of fine-grained visual details, while sequential inference introduces substantial latency. To address these limitations, we propose VLALight, a lightweight end-to-end vision-language-action framework that directly maps intersection observations and signal-phase information to discrete signal actions. To handle the multi-view nature of TSC, VLALight combines multiple directional camera views into a unified visual input and uses textual instructions to establish their correspondence with traffic movements and signal phases. This design enables direct action prediction with a compact 0.5 B-parameter model, without intermediate image-to-text descriptions or handcrafted traffic-state representations. Experiments show that VLALight delivers the best emergency-vehicle service of all compared methods, reducing pooled emergency waiting time by 21.1% over the cascaded VLMLight while running in real time on local hardware and generalizing to unseen intersection topologies and traffic-flow patterns.

Kemou Jiang, Maonan Wang, Xingchen Zou +6

School of Transportation Science and Engineering, Beihang University, Beijing, China · The Chinese University of Hong Kong (Shenzhen), Shenzhen, China · The Chinese University of Hong Kong, Hong Kong SAR, China +2

Traffic Signal Control Efficient VLA Models Vision-Language-Action Models

Ask PDF

May 28, 2026

ReasonLight: A Multimodal Foundation Model-Enhanced Reinforcement Learning Framework for Zero-Shot Traffic Signal Control

Reinforcement learning (RL) has shown promise in traffic signal control (TSC). However, its reliance on predefined states limits responsiveness to observable open-world events that are absent from training data. IoT-enabled intersections provide heterogeneous observations from roadside sensors and cameras, creating opportunities to improve RL adaptability to such events. To this end, we propose ReasonLight, a multimodal foundation model-enhanced RL framework for zero-shot TSC. ReasonLight integrates three sources of information: structured traffic measurements, multi-view camera observations, and candidate phase decisions from a pre-trained RL controller. Given an RL-proposed phase, ReasonLight extracts visual semantics from multi-view images and aligns them with compact sensor-derived scene descriptions. This alignment enables a semantic-guided refinement module to either preserve or adjust the proposed action according to traffic rules and event semantics. To ensure operational reliability, refined actions are constrained by the set of available phases. Any invalid decision is rejected, and the system falls back to the original RL action. We evaluate ReasonLight on two types of rare events not seen during RL training: emergency vehicle priority and temporary traffic regulation. Experimental results show that ReasonLight achieves zero-shot adaptation without retraining. It reduces emergency vehicle waiting time by up to 88.7% compared with the RL-only backbone while preserving comparable routine traffic performance.

Aoyu Pang, Maonan Wang, Yuejiao Xie +3

School of Science and Engineering, The Chinese University of Hong Kong, Shenzhen, China · Department of Mechanical and Automation Engineering, The Chinese University of Hong Kong, Hong Kong · Shanghai AI Laboratory, Shanghai, China +1

RL Control Intelligent Transportation Systems Traffic Signal Control Multimodal Foundation Models

Ask PDF

Abstract

Traffic Signal Control

Explore similar work

VLALight: A Vision-Language-Action Model for Traffic Signal Control

VLALight: Lightweight Vision-Language-Action Models for Emergency-Aware Traffic Signal Control

ReasonLight: A Multimodal Foundation Model-Enhanced Reinforcement Learning Framework for Zero-Shot Traffic Signal Control