acceptodds
Under review as a conference paper at ICLR 2027

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

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.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Traffic Signal Control, Sim-to-Real Transfer, Human Demonstration, Intelligent Transportation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_qBQdOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents BEVLight, a visual traffic signal control framework that directly maps UAV bird’s-eye-view observations to signal phases. The method grounds visual features onto lane-level representations, composes them through the junction topology into movement- and phase-level representations, and applies a shared scorer to rank feasible phases across heterogeneous intersection layouts and signal plans. The model is trained in two stages, first using simulator-provided traffic states to learn control-relevant lane and movement representations, and then using human demonstrations with full phase rankings to train the decision policy. The paper evaluates zero-shot generalization to unseen traffic demands, unseen signal plans, and unseen intersections, and additionally studies transfer from simulated observations to real UAV imagery.

Critical strengths
  1. Clear and practically meaningful problem formulation.
    The paper addresses a concrete deployment bottleneck in visual traffic signal control: existing methods often rely on structured traffic states or fixed intersection configurations, while BEVLight aims to directly operate on UAV visual observations and generalize across heterogeneous junctions and signal plans. The lane-to-movement-to-phase formulation provides a natural way to remove dependence on fixed action indices.

  2. Well-motivated topology-grounded architecture.
    The proposed architecture closely follows the structure of the control problem. Visual features are first grounded onto individual lanes, then composed into movement and phase representations according to the junction topology, and finally scored using a shared candidate-phase scorer. This design is intuitive and well aligned with the desired cross-layout generalization.

  3. Reasonably complete closed-loop evaluation.
    The paper evaluates progressively harder generalization settings, including unseen traffic demands, unseen signal plans, and jointly unseen intersections, plans, and demands. The reported results are competitive against both conventional and structured-state RL controllers, and the ablations provide evidence for the contributions of visual pretraining and ranking supervision. The paper also includes an offline simulation-to-real study on real UAV observations.

Critical weaknesses
  1. The novelty boundary relative to prior visual TSC and topology-generalizable TSC is not sufficiently clear.
    Several key ingredients of BEVLight have close precedents. Prior work has already studied direct visual traffic signal control from raw image or BEV observations, while recent generalizable TSC methods have used movement-based representations to unify heterogeneous intersection topologies and signal phases. Therefore, the main contribution appears to lie in the combination of UAV visual grounding, explicit lane-to-movement-to-phase composition, and human ranking supervision, rather than in visual TSC or topology-aware generalization individually. The paper would benefit from positioning this novelty more explicitly against the closest prior work.

  2. The baseline set does not fully support the paper’s strongest claims.
    The current visual baselines are mainly perception-to-structured-state pipelines, rather than competitive end-to-end visual TSC methods. This makes it difficult to determine whether the gains come specifically from BEVLight’s topology-grounded formulation or simply from learning control directly from visual observations. Similarly, since cross-topology generalization is a central claim, stronger comparison with recent universal/generalizable TSC methods would be important. In particular, a direct visual-control baseline and a movement-based cross-topology baseline would make the empirical contribution substantially more convincing.

Questions for authors

Can the authors strengthen the baseline comparison with (i) a direct visual TSC baseline, e.g., a TrafficDojo-style controller that uses the same UAV BEV observations but does not employ the proposed lane-to-movement-to-phase composition, and (ii) a recent movement-based generalizable TSC method such as Unicorn under the same unseen-plan and unseen-junction splits?

3 comments
AuthorsOpenReview ↗

Response to Reviewer qBQd

We sincerely thank you for the constructive feedback and positive assessment. In the revised manuscript, we have clarified the novelty relative to prior TSC methods and added new baseline comparisons with vision-only VLMLight variants in Appendix D.2, with all revisions highlighted in blue.

W1/Q1: Novelty Boundary and Direct Visual TSC

UAVs provide flexible, on-demand BEV observations that capture intersection geometry and lane-level traffic conditions in a unified view, making them well suited for visual TSC. BEVLight is designed to map raw BEV imagery directly to signal decisions across heterogeneous intersection layouts and signal plans, while learning across diverse visual styles to generalize to unseen traffic demands and real-world visual conditions. Unlike prior topology-generalizable controllers that rely on pre-estimated lane scalars or visual controllers tied to fixed action spaces, BEVLight decouples visual observations from fixed action indices and learns appearance-invariant phase rankings.

A TrafficDojo-style controller takes the raw image directly as input to predict discrete action indices, but it cannot be jointly trained or generalized across different configurations because the movement semantics of each action index are not reflected in the road image itself. As illustrated in Figure 4, for Plan A and Plan B of the same intersection such as Gaojiaoyuan, action index 1 represents opposing through movements in Plan A, whereas it represents single-approach through and left-turn movements in Plan B. Consequently, directly mapping BEV images to action indices without lane-to-movement-to-phase grounding creates conflicting action labels across plans during training and results in complete intersection gridlock on unseen signal plans and junctions, confirming the necessity of our topological grounding and composition.

The details of the novelty positioning and boundaries with prior work have been clarified in Abstract and Introduction of the revised manuscript.

W2/Q2: Additional Baseline Comparisons

Under the unseen-demand, unseen-plan, and unseen-junction splits, Table 1 compares BEVLight against representative RL-based TSC methods that receive ground-truth structured traffic states from the simulator, as well as two-stage visual baselines, since very few existing methods directly support generalizable control from UAV BEV imagery.

To further strengthen the baseline comparisons, we have added an evaluation of the recent vision-language controller VLMLight in Appendix D.2. For a fair comparison under purely visual inputs, we do not provide ground-truth structured lane queue lengths to VLMLight; instead, we evaluate both VLMLight and VLMLight-Think using DeepSeek-V4.1-Flash-1017 with reasoning mode enabled by providing the UAV BEV image combined with the visual mask diagrams of all candidate phases to select the target phase index. Table 8 in Appendix D.2 reports the closed-loop average vehicle waiting time in seconds:

Junction / PlanMaxPressureAttendLightVLMLightVLMLight-ThinkBEVLight
Guanghua A9.198.4813.9216.297.95
Guanghua B9.129.0234.5727.429.04
Songdo A13.3813.8715.5715.4912.77
Songdo B19.4719.16--18.10

Here, - indicates severe intersection gridlock where vehicles fail to clear. Without structured traffic states, general-purpose vision-language models struggle to aggregate fine-grained queues across lanes, particularly for opposing-movement phases in Plan B, leading to high waiting times on Guanghua B and gridlock on Songdo B. In contrast, BEVLight consistently outperforms both vision-only VLMLight variants, prevents gridlock on Songdo B, and matches or surpasses the oracle structured-state baselines.

The details of the additional baseline comparisons and domain generalization experiments have been clarified in Related Work and Appendix D.1 (Page 27) of the revised manuscript.

Reviewer_qBQdOpenReview ↗

Response to rebuttal

Thank you for your thorough and thoughtful rebuttal. I believe your responses and additional experiments have adequately resolved my concerns.

However, given my limited expertise in this specific task, I do not feel sufficiently confident to assign the highest score of 4. I appreciate your efforts and wish you all the best with the remainder of the rebuttal process.

Thank you so much for your positive feedback! We really appreciate your time and thoughtful comments throughout the review process. Your suggestions have been very helpful in improving our paper.

We wish you all the best!

Reviewer_HbzrOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

BEVLight maps UAV bird’s-eye-view observations to signal-phase scores through lane–movement–phase topology. Representations are pretrained using simulator traffic states, followed by policy learning from human phase selections and complete rankings. Generalization is evaluated across eight individual junctions, demand patterns, and signal plans, with additional offline predictions on real UAV recordings.

Critical strengths
  1. The topology-based scorer addresses a concrete transfer problem. Constructing phase representations from their constituent lanes and movements avoids relying on fixed action indices. This is a sensible design for sharing a policy across junctions and signal plans.
  2. The lightweight implementation is promising. The model has approximately 1.26 million parameters and reports 8.59 ms inference latency on an RTX 4090, alongside competitive waiting times in several simulation settings. This supports further investigation of the approach.
Critical weaknesses
  1. The reliability of human rankings and their alignment with the control objective are insufficiently validated. Phase ranking requires traffic-control expertise and an understanding of downstream action effects, yet ranking criteria and inter-annotator agreement are inadequately reported. Episode-level waiting-time filtering does not validate complete rankings at individual states. Since simulation allows comparisons of subsequent action outcomes, the choice of human preferences over outcome-based supervision requires stronger justification.
  2. The evaluation does not sufficiently support the performance and generalization claims. 1) Baseline coverage is incomplete. The visual comparisons mainly use author-constructed two-stage pipelines, without a direct comparison to an existing end-to-end visual controller. Comparisons with recent RL methods applicable to this task are also insufficient to establish an advantage over advanced alternatives. 2) Scenario coverage and realism are limited. The experiments cover eight individual junctions, without evaluating network-level coordination. The constructed demands are not adequately related to measured flows and turning distributions. 3) Reproducibility are insufficiently documented. BEVLight’s training hyperparameters, freezing settings between stages, and training time are incompletely reported.
Questions for authors
  1. The full model’s average waiting time on Songdo A is 13.08 seconds in Table 1 but 12.77 seconds in Table 2. What accounts for this discrepancy?
  2. What criteria guide human phase rankings, and how is agreement between annotators assessed?
  3. With inputs and training conditions held fixed, what gains come separately from the topology hierarchy and actual human rankings?
3 comments
AuthorsOpenReview ↗

Response to Reviewer Hbzr (1/2)

We sincerely thank you for the thorough review and detailed feedback. In the revised manuscript, we have verified and updated the experimental results, clarified our annotator training protocol and implementation details, and added new evaluations on visual domain generalization and vision-only VLMLight baselines. All revisions are highlighted in blue.

Q1: Table Discrepancy

Thank you for catching this discrepancy. Each intersection has five traffic demand patterns (three for training and two held-out for evaluation), and the entry for Songdo A in Table 1 inadvertently mixed the training and evaluation demand patterns when computing the average. We have corrected the entry in Table 1.

W1/Q2: Human Ranking Reliability

To ensure the quality and reliability of the human demonstrations, all participating researchers were trained on core TSC concepts prior to formal annotation, including phase-movement mappings, right-of-way conflicts, and phase-transition dynamics, and completed practice episodes on the teleoperation platform. As shown in Figure 9, hovering over a candidate action highlights the corresponding lanes and permitted movements directly on the BEV view, enabling annotators to assess spatial queue distributions and competing approach demands. Episode-level filtering against MaxPressure further screens out suboptimal trajectories, retaining only episodes that achieve lower average waiting time than MaxPressure.

Regarding outcome-based supervision from simulation, traffic signal decisions have long-term downstream effects on queue evolution. Evaluating subsequent action outcomes over a multi-step horizon H requires branching all K feasible phases at every decision step, resulting in an exponentially large search space of K^H simulator rollouts per state that is computationally intractable to traverse across thousands of decision states. As confirmed by our ablation study in Table 2, learning from human preference rankings consistently reduces average waiting time across all unseen scenarios.

The details of the annotator training and ranking rationale have been clarified in Appendix C.1 (Page 20 & 21) of the revised manuscript.

Q3: Gains from Pretraining and Human Rankings

With inputs and training conditions held fixed, the topology hierarchy and human rankings contribute distinct gains. First, the lane-to-movement-to-phase topology hierarchy is essential for cross-plan and cross-junction generalization: without it, mapping BEV images directly to discrete action indices fails when action semantics change across signal plans, as illustrated in Figure 4. Table 2 further confirms that traffic-aware pretraining is needed to effectively learn these lane and movement representations.

Second, holding this pretrained topology-grounded architecture fixed, Table 2 isolates the gain from actual human preference rankings over single-action labels. As shown in Figure 14(a), non-selected phases can differ substantially, with the first- and second-ranked phases serving active queues while the third-ranked phase has no waiting vehicles. Incorporating human rankings via the soft ranking loss consistently reduces waiting time across all four unseen scenarios, improving Guanghua A from 8.78 s to 7.95 s and Songdo A from 12.86 s to 12.77 s.

AuthorsOpenReview ↗

Response to Reviewer Hbzr (2/2)

W2: Experimental Coverage and Reproducibility

To comprehensively validate the performance and generalization of BEVLight, we have expanded our experimental evaluation and completed the implementation documentation. First, as shown in Figure 15 and Table 7 of Appendix D.1, we added a cross-domain generalization study comparing single-style and multi-style training across three unseen visual styles. Whereas single-style training suffers severe congestion or gridlock under heavy cast shadows in Eval-Golden, multi-style BEVLight maintains stable waiting times across all unseen styles. Table 7 reports the closed-loop average vehicle waiting time in seconds, where - indicates severe intersection gridlock:

Junction / PlanSingle Eval-SunnySingle Eval-EveningSingle Eval-GoldenMulti Eval-SunnyMulti Eval-EveningMulti Eval-Golden
Guanghua A9.189.2227.929.159.4710.28
Guanghua B9.9310.15-9.149.169.91
Songdo A13.5013.46-13.6314.5413.70
Songdo B15.3417.82-17.2318.0416.88

Second, in Appendix D.2, we added comparisons against two vision-only adaptations of VLMLight using DeepSeek-V4.1-Flash-1017, namely VLMLight and VLMLight-Think with reasoning mode enabled, alongside the oracle structured-state baselines MaxPressure and AttendLight. Both VLMLight variants receive BEV images and candidate phase masks without ground-truth lane scalars. The results show that BEVLight consistently outperforms both VLMLight variants, prevents gridlock on Songdo B, and matches or surpasses the oracle structured-state baselines. Table 8 reports the closed-loop average vehicle waiting time in seconds:

Junction / PlanMaxPressureAttendLightVLMLightVLMLight-ThinkBEVLight
Guanghua A9.198.4813.9216.297.95
Guanghua B9.129.0234.5727.429.04
Songdo A13.3813.8715.5715.4912.77
Songdo B19.4719.16--18.10

In addition, we clarified in Appendix B.3 how approach inflows and directional turning proportions across our eight real-world intersection layouts are configured to reflect typical urban arterial traffic, while noting that a single UAV viewpoint targets individual junction control and network-level coordination is left for future work. Finally, we documented the complete training hyperparameters and Stage I to Stage II module freezing settings in Appendix B.1 to ensure full reproducibility.

The details of the implementation settings, traffic demand configurations, and additional baseline and domain experiments have been clarified in Appendix B.1 (Page 16), Appendix B.3 (Page 18), Appendix D.1 (Page 27), and Appendix D.2 (Page 28) of the revised manuscript.

Dear Reviewer,

As the discussion period comes to a close, we thank you again for your constructive feedback. In our rebuttal and revised manuscript, we have thoroughly addressed all of your concerns through our clarifications on human ranking supervision, complete reproducibility documentation, and new closed-loop evaluations across unseen visual styles and VLMLight baselines.

Overall, BEVLight introduces a generalizable UAV-based visual TSC framework that decouples raw BEV observations from fixed action indices and learns appearance-invariant phase rankings across diverse visual styles, bridging simulated policy learning and zero-shot transfer to real-world UAV imagery. We sincerely hope that these revisions and additional results have fully resolved your concerns and can be reflected in your final evaluation.

Reviewer_JU8QOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

This paper focuses on BEV-based traffic signal control and proposes BEVLight for these scenarios. Specifically, a model is designed to extract the lane-level visual features and to build the phase-level representations. Human preference is introduced to enhance the control performance of BEVLight. Experimental results demonstrate that BEVLight can outperform most of the baselines.

Critical strengths
  1. This paper introduces new UAV-based traffic signal control scenarios and uses BEVs as input to control the signals.

  2. The design of the model structure is reasonable, taking the topology structure into account.

Critical weaknesses
  1. The main design of the model is somewhat similar to FRAP [1] and AttendLight [2].

  2. There is some inconsistency between the introduction and the experiment section, such as the structure of intersections and domain differences. Meanwhile, the baselines and scenarios are too simple and old-fashioned. More complex scenarios could be introduced to validate the performance of the method.

[1] Zheng, Guanjie, et al. "Learning phase competition for traffic signal control." Proceedings of the 28th ACM international conference on information and knowledge management. 2019.

[2] Oroojlooy, Afshin, et al. "Attendlight: Universal attention-based reinforcement learning model for traffic signal control." Advances in Neural Information Processing Systems 2020.

Questions for authors
  1. FRAP and AttendLight also build lane-level representations and use them to make the decision. It would be better for the authors to discuss the differences among these methods.

  2. In the introduction, different domains are introduced, such as various weather conditions and illumination. But in the experimental part, only phase predictions on real UAV imagery are provided without weather or illumination differences.

  3. The newest baseline introduced in the experimental part, UniTSA, was released in 2023. Therefore, more new baselines could be introduced to better understand the performance improvement of the proposed method, such as GESA [3] and VLMLight [4]. It could be okay if the proposed method can achieve comparable performance to VLMLight with better efficiency.

[3] Jiang, Haoyuan, et al. "A general scenario-agnostic reinforcement learning for traffic signal control." IEEE Transactions on Intelligent Transportation Systems 2024.

[4] Wang, Maonan, et al. "VLMLight: Safety-critical traffic signal control via vision-language meta-control and dual-branch reasoning architecture." Advances in Neural Information Processing Systems 2026.

4 comments
AuthorsOpenReview ↗

Response to Reviewer JU8Q (1/2)

We sincerely thank you for the constructive feedback. In the revised manuscript, we have clarified the distinctions from prior methods and added new experiments on visual domain generalization and VLMLight baselines, with all revisions highlighted in blue.

W1/Q1: Differences from Existing TSC Methods

UAVs provide flexible, on-demand BEV observations that capture intersection geometry and lane-level traffic conditions in a unified view, making them well suited for visual TSC. However, learning a generalizable visual policy across different intersection layouts and signal plans requires decoupling visual observations from fixed action indices, as the same action index can correspond to different traffic movements across configurations. BEVLight is specifically designed to address this challenge by learning a generalizable controller that maps raw BEV observations directly to signal decisions across heterogeneous layouts, signal plans, and visual conditions.

Existing methods such as FRAP [1] and AttendLight [2] assume direct access to structured lane-level states, such as queue lengths and vehicle counts, to make RL decisions. While inspired by their topological perspective, simply cascading a visual detector with such controllers cannot address the broader challenge of visual generalization across signal plans and visual domains, as perception errors and domain shifts propagate through the intermediate state-estimation bottleneck. Instead, BEVLight learns an end-to-end visual-topological policy directly from raw BEV imagery to score candidate phases across heterogeneous configurations, avoiding explicit detection or scalar reconstruction. In Table 1, BEVLight consistently outperforms both two-stage baselines, YOLO-C and BEV-S, which feed detected or estimated lane states into AttendLight, confirming the advantage of end-to-end visual-topological control. Furthermore, to improve training efficiency and data utilization, BEVLight learns relative phase priorities from human preference rankings over all feasible phases to adapt across varying traffic demands. To generalize across diverse visual conditions without collecting real-world control data, BEVLight trains across varied rendered visual styles to learn appearance-invariant representations that transfer zero-shot to real UAV imagery.

The details of the motivation and differences from prior topological TSC methods have been clarified in Abstract, Introduction, and Conclusion of the revised manuscript.


W2/Q2: Visual Domain Generalization and Intersection Structures

As shown in Figure 3 and Figures 10 to 13, our web-based demonstration platform records the underlying vehicle states alongside each human decision, allowing a single human annotation to be re-rendered into diverse visual styles. Training across varied weather and illumination styles is designed to learn appearance-invariant representations that generalize to unseen visual conditions and real UAV scenes.

To further evaluate zero-shot domain generalization, we added closed-loop experiments in Appendix D.1 comparing a single-style model trained only on Morning against multi-style BEVLight. Both models are tested on three unseen realistic styles illustrated in Figure 15: Eval-Sunny, Eval-Evening with streetlights, and Eval-Golden with heavy cast shadows across the lanes. Table 7 reports the closed-loop average vehicle waiting time in seconds:

Junction / PlanSingle Eval-SunnySingle Eval-EveningSingle Eval-GoldenMulti Eval-SunnyMulti Eval-EveningMulti Eval-Golden
Guanghua A9.189.2227.929.159.4710.28
Guanghua B9.9310.15-9.149.169.91
Songdo A13.5013.46-13.6314.5413.70
Songdo B15.3417.82-17.2318.0416.88

Here, - indicates severe intersection gridlock where vehicles fail to clear. While both models perform comparably under uniform lighting in Eval-Sunny and Eval-Evening, the single-style model suffers severe delay or gridlock under heavy cast shadows in Eval-Golden, whereas multi-style training maintains stable performance across all three unseen styles. Regarding intersection structures, our real-world benchmark already covers three-arm T-junctions such as Beishahe and Songdo T, skewed layouts such as Zhijingdao, and four-arm intersections with heterogeneous lane topologies and signal plans.

The details of the intersection structures and visual domain evaluations have been clarified in Introduction and Appendix D.1 (Page 27) of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer JU8Q (2/2)

Q3: Comparison with VLMLight and Recent Baselines

Our goal is to make signal control decisions purely from UAV BEV imagery without external traffic-state sensors. In contrast, GESA [3] operates on structured traffic states rather than raw visual observations, and while VLMLight [4] incorporates visual inputs, its dual-branch design relies on structured traffic states such as queue lengths for routine RL control while using the VLM branch for safety-critical events, rather than operating solely on images. We have added the citation and discussion of GESA [3] in Related Work.

For a fair comparison under purely visual inputs, we have added an experiment in Appendix D.2 evaluating two vision-only adaptations of VLMLight using DeepSeek-V4.1-Flash-1017, namely VLMLight and VLMLight-Think with reasoning mode enabled, where the model receives the UAV BEV image and the visual mask diagrams of all candidate phases to select the target phase without ground-truth structured states. Table 8 in Appendix D.2 reports the closed-loop average vehicle waiting time in seconds:

Junction / PlanMaxPressureAttendLightVLMLightVLMLight-ThinkBEVLight
Guanghua A9.198.4813.9216.297.95
Guanghua B9.129.0234.5727.429.04
Songdo A13.3813.8715.5715.4912.77
Songdo B19.4719.16--18.10

Here, - indicates severe intersection gridlock where vehicles fail to clear. Without queue lengths, general-purpose vision-language models struggle to aggregate lane-level queues across opposing approaches in Plan B, causing high delay on Guanghua B and gridlock on Songdo B even with reasoning mode enabled. In contrast, BEVLight consistently outperforms both VLMLight variants, prevents gridlock on Songdo B, and matches or surpasses the structured-state baselines using only 1.26M parameters and 8.59 ms inference latency per step.

The details of the discussions and experimental comparisons with recent baselines have been clarified in Appendix D.2 (Page 28) of the revised manuscript.

Thank you for the response and additional experiments. The lighting experiments address part of my concerns, but I am maintaining my original score because the methodological novelty remains unclear.

FRAP and AttendLight already construct phase representations from lane- or movement-level features. The response clarifies differences in inputs and supervision, but does not sufficiently establish the contribution of BEVLight’s topology-based composition and scoring.

Thank you for taking the time to read our response and for confirming that the additional domain experiments address your concerns regarding lighting variations. We would like to clarify the core research goal and methodological contributions of our work.

The goal of this paper is to enable generalizable visual traffic signal control directly from UAV BEV observations, as UAVs provide flexible, on-demand monitoring without permanent roadside infrastructure and capture intersection geometry and lane-level traffic conditions in a unified view. To learn a single UAV-based visual controller that generalizes zero-shot across different intersection layouts, signal plans, and unseen traffic demands, our work makes two core methodological contributions:

  1. Decoupling BEV visual observations from fixed action indices across layouts and signal plans: In raw BEV imagery, the visual observation contains only road geometry and vehicles without explicit action semantics, while the same discrete action index corresponds to completely different movements across signal plans and intersections. As shown in Table 8, existing visual control paradigms and vision-language models such as VLMLight struggle or suffer complete gridlock when movements are paired across opposing approaches. BEVLight resolves this visual-to-action ambiguity by grounding 2D BEV features onto the road geometry and composing them according to the active signal plan, enabling a shared visual policy to score candidate phases across heterogeneous layouts and phase configurations without intermediate object detection or scalar state reconstruction, outperforming two-stage visual pipelines by approximately 10% in Table 1.

  2. Generalizing across unseen traffic demands and real-world visual domains: To prioritize competing phases under varying congestion patterns and bridge the visual gap between simulation and real UAV imagery without real-world control data collection, BEVLight learns relative phase priorities from human preference rankings over all feasible phases while aligning visual representations with appearance-invariant traffic semantics across diverse visual styles. As confirmed by our cross-domain evaluation in Table 7 and offline real-UAV study, this paradigm enables robust zero-shot generalization across severe visual shifts such as heavy cast shadows in Eval-Golden and direct transfer to real-world UAV videos.

We hope this clarifies our core contribution: BEVLight aims to advance a more flexible, infrastructure-free paradigm for traffic signal control, bridging the long-standing gap between simulated policy learning and direct evaluation on real-world visual observations.

Comments on the paper

2 comments

Dear Area Chairs,

I hope you're doing well.

We are the authors of Submission #39462. We have provided detailed responses to all reviewers' comments and made corresponding revisions to the manuscript. So far, only Reviewer qBQd has responded. We would greatly appreciate it if the remaining reviewers could engage in further discussion based on our responses, which would help us further improve the quality of our paper.

Thank you for your time and support.

Best regards, Authors of Submission #39462

Dear Area Chairs,

We would like to provide a brief summary of the core contributions of our work and the key improvements made in the revised manuscript to assist in your final evaluation:

  1. A Flexible, Generalizable UAV BEV Paradigm for Visual TSC: Existing TSC methods mostly rely on pre-estimated structured traffic states or fixed roadside sensors tied to specific junctions. We introduce BEVLight to enable on-demand visual signal control directly from UAV bird's-eye-view observations, decoupling raw visual inputs from fixed action indices so that a single policy can operate across different intersection layouts and signal plans without retraining.
  2. Appearance-Invariant Human Preference Learning and Teleoperation Platform: To prioritize competing phases under varying traffic demands and bridge the simulation-to-real visual gap without costly real-world control data collection, we develop a state-preserving virtual teleoperation platform and train BEVLight from full human phase rankings across diverse visual styles, aligning visual features with appearance-invariant traffic semantics.
  3. Comprehensive Zero-Shot Generalization and Sim-to-Real Validation: Closed-loop evaluations demonstrate that BEVLight generalizes zero-shot across unseen demands, signal plans, intersections, and challenging lighting conditions with cast shadows.

We have also thoroughly updated the manuscript to incorporate all reviewer suggestions. Thank you very much for your time and effort in handling our submission.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.