acceptodds
Desk rejected at ICLR 2027

CoopLight: Cooperative Vision-Language Agents for Long-Tail Traffic Control

Desk rejected at ICLR 2027, so left out of the paper lists and the map.

Abstract

Traffic light signals and connected and automated vehicles (CAVs) are increasingly interdependent at urban intersections, yet they are still predominantly studied in isolation. Existing methods mainly rely on low-dimensional traffic states, limiting their ability to handle long-tail scenarios such as incidents, jaywalking pedestrians, and emergency-priority events. Such settings require semantic scene understanding, context-aware decision making, and timely cooperation between infrastructure-side and vehicle-side agents. To address these challenges, we introduce CoopLight, a vision-language framework for traffic signal and CAV control. Agent-specific visual question answering (VQA) tasks ground spatial scene understanding and decision making in multi-view observations. Multi-agent co-training further enhances agents’ ability to integrate local observations with received semantic reports. This learned cooperative reasoning enables signal and vehicle agents to use complementary evidence and adapt their decisions under partial observability. Extensive experiments on two real signalized intersections under multiple long-tail traffic scenarios show that CoopLight consistently improves traffic efficiency over rule-based, reinforcement-learning-based, and prior vision-language-model-driven baselines, demonstrating its effectiveness and proactivity for next-generation urban mobility.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Multi-Agent Decision Making, Vision-Language Models, Traffic Signal Control

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_wzdKOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces CoopLight, a vision-language framework for cooperative traffic signal and connected autonomous vehicle control under long-tail traffic scenarios. It first fine-tunes agents on perception and decision-making VQA tasks, then uses GRPO-based multi-agent co-training to improve semantic report accuracy and cooperative action selection. The agents exchange event locations and confidence levels to integrate complementary observations. Experiments on simulated replicas of two real-world intersections show reduced waiting and travel times, with additional evaluations supporting cross-intersection transfer and robustness to adverse visual conditions.

Critical strengths
  1. Well-motivated: The paper addresses practically important long-tail traffic events and clearly explains why compact numerical states are insufficient for effective control.

  2. The authors evaluate actual waiting and travel times in traffic simulation, providing evidence of control effectiveness beyond perception accuracy.

  3. Experiments cover multiple event types, co-occurring events, cross-intersection transfer, and adverse visual conditions, while ablations support the benefits of vehicle control and communication.

  4. The framework is well suited to cooperative traffic control, where roadside and vehicle agents provide complementary observations that support informed signal decisions and proactive vehicle maneuvers.

Critical weaknesses
  1. The experiments demonstrate the benefits of vehicle control and communication, but do not isolate the contributions of multi-agent co-training, confidence labels, or self-reflection. The necessity of these components remains unclear.

  2. Evaluation is limited to simulated environment of two intersections and predefined event categories. Although transfer and visual robustness experiments are included, generalization to diverse real-world traffic conditions remains unverified.

Questions for authors
  1. Could the authors provide controlled ablations isolating the effects of GRPO-based multi-agent co-training, confidence labels, and self-reflection? In particular, how does the full method compare with agents trained only through supervised fine-tuning while retaining the same communication mechanism?
  2. Could the authors clarify whether CoopLight generalizes to unseen event types or configurations beyond the predefined training scenarios? Additional evaluation on more intersection layouts or held-out event configurations would help establish the scope of its generalization.
4 comments

We thank the reviewer for the positive assessment of the long-tail traffic-control setting, the closed-loop evaluation, and the role of complementary TLS and CAV observations. We appreciate the request to isolate the contributions of multi-agent GRPO co-training, communicated confidence, and self-reflection, and to examine generalization beyond the original scenarios and intersection layouts. Below, we provide controlled component ablations and evaluations on held-out event categories, event combinations, and additional junction layouts, while distinguishing simulation-based transfer from real-world validation and network-wide coordination.

Response to Weakness 1 & Question 1

Thank you for this question. The manuscript's Table 1 separates vehicle-side control (CoopLight w/o CAV) from semantic communication (CoopLight w/o Comm). We further ablate multi-agent GRPO co-training, the confidence field in communicated reports, and the self-reflection step. All variants use the same evaluation scenarios, and each table entry reports Overall AWT / CAV AWT in seconds.

ScenarioIntersectionCoopLight w/o Commw/o Co-training (SFT + Comm)w/o Report Confidencew/o Self-reflectionCoopLight
JaywalkingCAL22.1 / 11.620.6 / 10.520.3 / 10.619.6 / 9.818.9 / 9.4
JaywalkingYMT23.5 / 19.821.6 / 14.321.9 / 15.120.3 / 11.219.5 / 10.2
IncidentCAL16.2 / 5.114.9 / 4.814.6 / 4.715.1 / 4.913.3 / 4.2
IncidentYMT15.9 / 15.314.6 / 14.814.8 / 14.914.9 / 14.813.2 / 14.1
BarrierCAL24.5 / 3.722.3 / 3.721.8 / 3.822.6 / 3.719.6 / 3.6
BarrierYMT27.2 / 13.525.4 / 13.025.6 / 13.125.9 / 13.223.2 / 12.3

With the communication mechanism retained, SFT-only agents achieve mean Overall AWT of 19.90 s, versus 21.57 s for CoopLight w/o Comm and 17.95 s for full CoopLight. Relative to SFT + communication, joint GRPO reduces Overall AWT by 9.8%. Agent-specific SFT (Eq. 1) trains perception and decision making, while cooperative SFT (Eq. 2) teaches receivers to integrate local observations with ground-truth semantic reports. Multi-agent GRPO (Eq. 3) further optimizes report accuracy and action correctness using agents' generated reports in joint rollouts. This comparison identifies the benefit of co-training beyond supervised initialization, while recognizing that GRPO changes both reporting and decision making.

Removing the confidence field from communicated reports, so that messages contain only event type and location, increases Overall AWT by 10.5% compared with full CoopLight. The increase is larger at YMT (11.4%) than CAL (9.5%), consistent with the more frequent occlusions of its narrower approaches. This specifically tests the value of report-level confidence, not the removal of all confidence-aware perception training.

Without self-reflection on the outcome of the previous phase or maneuver, Overall AWT increases by 13.3% on incident and barrier scenarios, versus 3.9% on jaywalking. Persistent disruptions give the controller more opportunities to repeat an ineffective response, whereas pedestrian crossings tend to be transient. Each ablation degrades performance across the evaluated settings, supporting the contribution of these components within the proposed framework.

Response to Weakness 2 & Question 2

CoopLight reasons about the traffic consequences of observed events, including blocked movements, remaining dischargeable demand, and emergency priority. The phase-to-movement mapping and camera approach directions are specified for each intersection, allowing the agents to interpret new layouts without assuming the same phase indices or driving conventions.

The main experiments already use disjoint training and evaluation routes within each event category and vary event placement and traffic density. Barriers include construction cones and lane restrictions across different approaches and lanes. Six additional co-occurring-event routes evaluate combinations of long-tail events, where CoopLight obtains lower Overall and CAV AWT than CoTV (manuscript Figure 4a). These results concern unseen configurations and combinations of known event categories, rather than completely unconstrained real-world anomalies.

We further hold out one scenario category at a time from task-specific fine-tuning, train on the remaining categories, and evaluate on the held-out category. Entries report Overall AWT / CAV AWT (seconds).

Held-out typeIntersectionVLMLight (in-domain reference)CoopLight (held-out type)CoopLight (in-domain)
JaywalkingCAL26.8 / 27.522.6 / 11.818.9 / 9.4
JaywalkingYMT28.3 / 28.723.4 / 13.619.5 / 10.2
IncidentCAL27.1 / 29.714.8 / 4.913.3 / 4.2
IncidentYMT26.4 / 27.914.6 / 15.013.2 / 14.1
BarrierCAL28.9 / 37.321.9 / 4.319.6 / 3.6
BarrierYMT30.8 / 36.525.8 / 13.423.2 / 12.3

Holding out incidents and barriers increases CoopLight's Overall AWT by 10.9% and 11.4%, respectively, relative to its in-domain model; holding out jaywalking yields a larger 19.8% increase. The held-out CoopLight results are also below the in-domain VLMLight reference in all six comparisons, but that comparison is contextual rather than a matched held-out baseline. Here, an event type is unseen during task-specific fine-tuning, not necessarily unknown to the pretrained VLM.

We additionally evaluate two isolated intersections derived from the InTAS Ingolstadt [1] and TAPAS Cologne [2] SUMO urban-mobility scenarios, which provide city-derived road networks and traffic-demand specifications. Long-tail events are introduced on different approaches and lanes, and the policies trained on CAL and YMT are evaluated without fine-tuning on these additional layouts. These tests assess transfer to new junctions, not network-wide coordinated signal control. Entries report Overall AWT / CAV AWT (seconds).

IntersectionScenarioMax-PressureVLMLightCoopLight w/o CommCoopLight
IngolstadtJaywalking35.2 / 22.833.9 / 30.225.8 / 15.921.7 / 11.8
IngolstadtIncident30.4 / 24.131.2 / 31.518.7 / 9.615.6 / 7.9
IngolstadtBarrier36.8 / 25.634.7 / 37.928.9 / 10.423.9 / 9.1
CologneJaywalking33.6 / 21.532.4 / 29.124.6 / 14.720.8 / 11.3
CologneIncident29.1 / 22.830.3 / 30.817.9 / 8.815.1 / 7.4
CologneBarrier35.4 / 24.733.8 / 36.227.6 / 9.722.8 / 8.6

On these additional junctions, CoopLight averages 20.0 s Overall AWT, compared with 32.7 s for VLMLight and 23.9 s without communication, corresponding to reductions of 38.9% and 16.4%, respectively. The results complement the original CAL–YMT transfer and weather/illumination evaluations, but do not establish transfer to arbitrary cities or validated real-world traffic operation. Scaling semantic cooperation across connected intersections and evaluating field-calibrated traffic remain important future directions.

References

[1] S. Lobo, S. Neumeier, E. M. G. Fernandez, and C. Facchi. InTAS—The Ingolstadt Traffic Scenario for SUMO. SUMO Conference Proceedings, 1:73–92, 2020. doi: 10.52825/scp.v1i.102.

[2] S. Uppoor, O. Trullols-Cruces, M. Fiore, and J. M. Barcelo-Ordinas. Generation and Analysis of a Large-Scale Urban Vehicular Mobility Dataset. IEEE Transactions on Mobile Computing, 13(5):1061–1075, 2014. doi: 10.1109/TMC.2013.27.

Thanks for the detailed relpy. The authors have addressed my concern. They added controlled experiments and discuss about the generalization of the method. I will keep my positive socre.

Thank you for your positive feedback. We sincerely appreciate your constructive suggestions, which have helped us strengthen our evaluation and clarify the generalization capabilities of our method. Thank you again for your time and encouragement.

Reviewer_2imxOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes CoopLight, a framework that uses semantic information exchange between traffic signal and vehicle agents to handle traffic events such as accidents, road obstacles, crossing pedestrians, and emergency vehicles. The traffic signal agent uses roadside multi-view images and traffic detector data to select signal phases, while each connected and automated vehicle (CAV) uses surround-view images, its own state, and received reports to select lateral maneuvers. Training combines supervision through perception and decision-making VQA tasks, cooperative supervised fine-tuning (SFT), and multi-agent GRPO with a shared reward.

The paper reports improvements in traffic efficiency on simulated replicas of two real-world intersections and includes evaluations of cross-intersection transfer, robustness to weather variations, and perception quality. My main concern is that the scope of control and the information available to each agent are not sufficiently specified, while the existing comparisons do not clearly disentangle the contributions of additional observations, differences in control authority, and the joint training procedure.

Critical strengths
  1. The problem has practical relevance. Roadside and onboard views can provide complementary evidence for coordinating signal phases and vehicle maneuvers. This is a promising approach to handling local occlusions, blocked lanes, and emergency vehicle prioritization, which can be difficult for either roadside or vehicle controllers to address independently. In particular, the distinction between ordinary queues and obstructed queues that cannot discharge has clear implications for effective traffic control.

  2. The complete system demonstrates reasonably consistent closed-loop improvements. Table 1 shows that CoopLight outperforms the reported baselines across most scenarios and metrics. The comparison with the variant without communication also supports the practical value of semantic information sharing. These results provide encouraging evidence for the effectiveness of the overall framework.

Critical weaknesses

W1. Unclear Control Execution and Safety.

The proposed framework focuses on high-level signal phases and lateral vehicle maneuvers, but the execution of these decisions remains insufficiently specified. Although Appendix B indicates that SUMO handles vehicle dynamics, it is unclear how speed control, braking, and lane-change trajectories are implemented. Moreover, the framework does not adequately explain how lane markings, road restrictions, and available gaps constrain maneuver selection, or how infeasible commands are handled. Since the current evaluation primarily reports traffic efficiency, the absence of safety-related metrics and simulator intervention statistics limits the assessment of the framework's practical feasibility.

W2. Unclear Cooperation and Information Assumptions.

While CoopLight introduces semantic communication between TLS and CAV agents, the communication and information-sharing mechanisms require further clarification. In particular, it remains unclear how reports from multiple CAVs are selected and integrated, how redundant or outdated information is handled, and what information each agent can access during execution. Although the paper distinguishes privileged expert supervision from agent observations, a clearer specification of execution-time information boundaries is necessary. Furthermore, the evaluation focuses on jointly trained agents, leaving the framework's applicability to independently trained or fixed-policy vehicles uncertain.

Questions for authors

Q1. Control Execution and Safety.

How are the CAV agents' high-level lane decisions translated into executable vehicle actions in SUMO? Could the authors clarify how illegal or unsafe lane changes are prevented and whether SUMO overrides infeasible commands? Reporting collision-related metrics or safety intervention frequencies would help establish the scope of the proposed control framework.

Q2. Multi-Agent Communication.

Section 4.3 defines semantic reports containing event types, locations, and confidence levels. How does the TLS agent integrate reports from multiple CAVs, particularly when information is redundant, conflicting, or outdated? Please also clarify the execution-time observations and communicated information available to each agent, and whether the intended deployment assumes jointly trained vehicle policies.

Q3. Comparison with Recent Methods.

Although VLMLight and CoTV provide relevant baselines, comparisons with recent approaches such as Traffic-R1 [1] and CoLLMLight [2] would further contextualize the contribution. Could the authors discuss their applicability to the proposed setting and provide comparisons where feasible, accounting for differences in observations, communication structures, and control authority?

Reference:

[1] Zou X, Yang Y, Chen Z, et al. Traffic-r1: Reinforced llms bring human-like reasoning to traffic signal control systems[C]//Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2026: 21823-21838.

[2] Yuan Z, Lai S, Liu H. Collmlight: Cooperative large language model agents for network-wide traffic signal control[C]//International Conference on Learning Representations. 2026, 2026: 66687-66715.

3 comments

We sincerely thank the reviewer for the thoughtful assessment and for recognizing the practical value of complementary roadside and onboard observations, particularly in distinguishing obstructed queues from routine congestion. We appreciate the concerns about clearer execution-time control and information boundaries, a more explicit account of report aggregation, and comparisons that account for differences in observation modalities and control authority. In response, we clarify how SUMO handles high-level vehicle commands, report safety and communication-robustness evaluations, examine fixed-policy cooperation, and compare against adapted Traffic-R1 and CoLLMLight signal controllers.

Response to Weakness 1 & Question 1

Thank you for raising this question. CoopLight produces high-level decisions, while SUMO handles vehicle motion. The TLS selects from the intersection's configured signal phases, while each CAV selects among available lateral maneuvers based on lane accessibility, road connectivity, and permitted movements. Lane-change requests are submitted through TraCI's changeLane interface and processed by SUMO's LC2013 model, which checks safe gaps to the target-lane leader and follower and allows cooperative speed adjustments. Infeasible requests are not executed. Longitudinal speed control and braking follow SUMO's Krauss car-following model, subject to speed limits and traffic signals. We evaluate maneuver feasibility through blocked lane-change attempts alongside traffic efficiency under the SUMO execution interface.

Table 1. Intervention statistics under the execution interface, averaged across jaywalking, incident, and barrier scenarios at both intersections (seconds for AWT/ATT).

MethodOverall AWT / ATT ↓CAV AWT / ATT ↓Blocked lane changes (%) ↓
CoTV28.32 / 51.9515.17 / 39.8738.9
CoopLight w/ SFT CAV w/o Comm22.45 / 47.9612.38 / 36.8431.6
CoopLight w/ SFT CAV w/ Comm19.90 / 46.0510.18 / 35.1024.8
CoopLight w/o Comm21.57 / 47.2811.50 / 35.9727.3
CoopLight17.95 / 44.288.97 / 33.8818.5

As shown in Table 1, CoopLight achieves the lowest waiting times and blocked lane-change rate. Communication reduces blocked requests even under SFT, while co-training brings further improvements. These trends are consistent with earlier anticipation of affected lanes and fewer infeasible maneuver requests.

We further evaluate collision and time-to-collision (TTC) metrics with safety checks disabled to examine the safety of executed maneuvers. TTC conflicts are defined by a minimum TTC below 3 s. Collision and conflict frequencies are normalized per 100 executed CAV lane changes.

Table 2. Safety evaluation with safety checks disabled, averaged across three long-tail scenarios at both intersections.

MethodOverall AWT / ATT ↓CAV AWT / ATT ↓Vehicle–Vehicle Collisions ↓Vehicle–Pedestrian Collisions ↓TTC Conflicts ↓Trip Completion (%) ↑
CoTV27.36 / 46.2219.85 / 39.472.500.0018.286.7
CoopLight w/ SFT CAV w/o Comm22.71 / 48.1512.69 / 37.101.670.3316.193.3
CoopLight w/ SFT CAV w/ Comm20.14 / 46.3010.47 / 35.321.250.0015.695.6
CoopLight w/o Comm21.83 / 47.5211.84 / 36.211.170.1715.895.6
CoopLight18.21 / 44.499.26 / 34.070.700.0015.297.8

As shown in Table 2, CoopLight achieves the lowest vehicle–vehicle collision and TTC-conflict rates, together with the highest trip completion rate. Its vehicle–vehicle collision frequency is 72.0% lower than CoTV's. Within the CoopLight variants, communication and co-training are associated with fewer unsafe maneuvers. These results suggest that the traffic-efficiency gains are accompanied by improved maneuver feasibility and safety outcomes in the evaluated simulations. We will incorporate the details and safety evaluations into the revision.

Response to Weakness 2 & Question 2

Thank you for these questions. Each agent reports the event type, affected location, and confidence level. CAVs transmit events with confidence of at least LIKELY. The TLS aggregates reports of the same event, retaining corroboration counts and timestamps. Reports expire after 8 s for static incidents or 4 s for moving emergency vehicles. Conflicting reports are reconciled using confidence, recency, corroboration, and local observations. Sender IDs and timestamps serve as communication metadata.

Cooperative SFT teaches receivers to integrate local observations and shared reports (Eq. 2), while multi-agent GRPO jointly rewards report consistency and action correctness (Eq. 3). This supports consensus, complementary reasoning, and conflict resolution without requiring receivers to accept every report.

We compare closed-loop performance using predicted and ground-truth reports containing event types and locations observable to each sender.

Table 3. Closed-loop performance with predicted and ground-truth event reports, averaged over three scenarios and two intersections (seconds).

MethodReport contentOverall AWT ↓Overall ATT ↓CAV AWT ↓CAV ATT ↓
SFT + communicationPredicted19.9046.0510.1835.10
SFT + communicationGround-truth18.1244.628.8633.71
CoopLightPredicted17.9544.288.9733.88
CoopLightGround-truth17.2143.808.3533.42

Co-training reduces the performance gap between predicted and ground-truth reports. Even with ground-truth event information, Overall AWT decreases from 18.12 s to 17.21 s, indicating benefits beyond report accuracy alone.

We evaluate decision accuracy under threefold report corroboration, 30% corruption of event or location claims, and a 3 s report delay, keeping local observations unchanged.

Table 4. Open-loop decision accuracy under report perturbations (%).

Report conditionTLS: SFT only ↑TLS: Co-trained ↑CAV: SFT only ↑CAV: Co-trained ↑
Original predicted reports78.484.681.286.9
Redundant reports: 3× corroboration78.184.480.986.7
Conflicting reports: 30% corrupted63.576.266.879.4
Outdated reports: 3 s delay67.978.870.381.5

Redundant reports have little effect on accuracy, while co-trained agents are more robust to conflicting and delayed reports than SFT-only agents.

During execution, the TLS policy observes four directional camera views, per-movement detector counts, and the phase-to-movement mapping defined by the intersection's signal program, together with the aggregated semantic reports it receives from CAVs. A CAV policy observes six surrounding-view images, its own speed and current lane, permitted maneuver options, and relevant local lane/route metadata, together with TLS reports concerning its approach. The policies do not receive ground-truth event labels, other agents' raw images, or privileged global vehicle states; simulator state is used to construct expert supervision and training rewards, not as policy input at test time.

To evaluate independently trained and fixed-policy vehicles, we compare SFT-only TLS and CAV policies operating together without GRPO co-training, and TLS-only GRPO with a fixed SFT CAV partner, against full joint co-training.

Table 5. Performance with separately updated or fixed compatible vehicle policies (seconds).

ConfigurationOverall AWT / ATT ↓CAV AWT / ATT ↓
CoopLight w/o Comm21.57 / 47.2811.50 / 35.97
SFT + communication19.90 / 46.0510.18 / 35.10
TLS GRPO + frozen SFT CAV18.86 / 45.129.95 / 34.71
CoopLight: joint co-training17.95 / 44.288.97 / 33.88

SFT-based agents can communicate without joint GRPO, and TLS optimization with a frozen CAV policy also improves performance. Joint co-training further reduces Overall AWT by 9.8% relative to SFT + communication and 4.8% relative to TLS-only GRPO. These results support cooperation with fixed-policy CAVs and the additional benefit of joint optimization. We will incorporate these details and evaluations into the revised manuscript.

Response to Question 3

Thank you for highlighting these relevant works. Traffic-R1 [1] performs reinforcement-trained reasoning for traffic signal control and supports communication across intersections, while CoLLMLight [2] develops asynchronous cooperative reasoning for network-wide signal control. Both rely primarily on structured and textual traffic inputs and select signal-control actions, rather than processing TLS/CAV camera images jointly or controlling CAV maneuvers. CoopLight instead combines visual evidence from infrastructure and vehicles, semantic TLS–CAV communication, and two distinct agent roles. We evaluate adaptations of Traffic-R1 and CoLLMLight as TLS-only baselines. Their structured inputs include traffic measurements and the intersection's phase-to-movement mapping. For event information, a classifier-based perception module with supervised event-classification and affected-lane-localization heads, trained on the corresponding visual data split, produces event information from camera observations.

Table 6. Adapted signal-controller comparison across three long-tail scenarios and two intersections (seconds).

MethodOverall AWT ↓Overall ATT ↓CAV AWT ↓CAV ATT ↓
Traffic-R1 (adapted)26.1248.9516.8440.12
CoLLMLight (adapted)27.0349.4017.6240.71
CoopLight w/o CAV24.1850.2720.3747.97
CoopLight17.9544.288.9733.88

CoopLight w/o CAV achieves lower Overall AWT than both adapted baselines, although it does not outperform them across all metrics. Full CoopLight achieves the best performance across all four metrics, highlighting the additional benefits of vehicle-side control and TLS–CAV semantic communication. We will incorporate these comparisons into the revised manuscript.

References

[1] Zou et al., Traffic-R1: Reinforced LLMs Bring Human-Like Reasoning to Traffic Signal Control Systems, ACL 2026.

[2] Yuan et al., CoLLMLight: Cooperative Large Language Model Agents for Network-Wide Traffic Signal Control, ICLR 2026.

Reviewer_Gwa2OpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CoopLight is a vision-language framework for coordinating traffic signals and autonomous vehicles. Signal agents analyze traffic camera images, and vehicle agents use egocentric views. They both identify events and share analysis. The signal agents observe and select signal phases, and vehicle agents change moving directions to avoid incidents. The method combines supervised fine-tuning and joint multi-agent training to formulate accurate reports and effective decisions. Evaluations in simulations show that their multi-agent cooperation can improve traffic efficiency. Their methods are helpful in scenarios with incidents, obstructions, and emergency vehicles, where long-tail features or events can be missed by previous methods with only local observations.

Critical strengths
  • The paper addresses a gap in previous methods that only relies on local observations. They bring signal and vehicle direction control into a single framework for long-tail intersection events.

  • Its cross-view mechanism offers a way to share observations across devices. They successfully improved the long-tail event detection abilities and coordination abilities.

Critical weaknesses
  • Agents share events, locations, and their confidence. However, they choose actions in separate action spaces without explicit coordination mechanisms. The paper does not establish an explicit coordination method for how a particular vehicle and traffic light could cooperate. The ablation on the communication mechanism only supports the effects of information sharing, but does not necessarily connect to the cooperation between vehicles and lights.

  • Their experiments use simulated replicas of only two intersections. Its effectiveness in region-level and city-level traffic networks remains a concern. Most previous methods have already focused on multi-intersection scenarios. How their method benefits multiple subsequent intersections, large-scale networks, and real-world traffic scenarios remains an important problem.

Questions for authors
  • Could you clarify how joint training coordinates traffic lights with vehicles, and isolate the benefit of this coordination from the improvement in traffic condition detection?

  • Could you report additional evidence supporting simulation fidelity and transferability on large-scale road networks?

  • Could you provide safety evaluations to show that the traffic efficiency gains do not compromise safety?

4 comments

We sincerely thank the reviewer for the constructive feedback and for recognizing the value of combining infrastructure-side and vehicle-side observations for long-tail traffic understanding and cooperative control. We appreciate the concerns about coordination across heterogeneous action spaces, distinguishing joint optimization from improved event detection, and simulation-based evaluation. Below, we clarify the multi-agent co-training objective and its ablations, provide additional evidence on simulation settings and cross-intersection transfer, and report safety and maneuver-intervention analyses.

Response to Weakness 1 & Question 1

Thank you for raising this question. CoopLight coordinates TLS and CAV agents through semantic communication and joint policy optimization, rather than explicit action-level commands or vehicle-specific signal reservations. Different action spaces do not preclude cooperative control: for example, CoTV (Guo et al., 2023) combines discrete signal decisions and continuous CAV acceleration control through state exchange rather than sharing planned actions. In CoopLight, cooperation consists of complementary decisions grounded in shared semantic information. When an incident blocks a lane, the TLS adjusts signal phases according to dischargeable demand while CAVs use shared reports to anticipate the obstruction and change lanes. During emergency events, the TLS prioritizes the relevant movement while CAVs select maneuvers that facilitate passage.

As described in Section 4.3, agent-specific SFT first teaches perception and decision making, and cooperative SFT further teaches each receiver to condition its decisions on local observations and communicated reports. During multi-agent GRPO co-training, TLS and CAV agents generate reports and actions in joint rollouts. Each rollout receives a shared reward that sums report consistency in event type and location and agent-specific action correctness (Eq. 3). Group-relative advantages are applied to both agents' generated tokens, encouraging both reporting and decision policies to adapt within the communicating system. We compare four variants with the same communication protocol and cooperative SFT initialization. SFT + communication performs no subsequent GRPO updates. TLS-only GRPO updates the TLS policy while holding the SFT CAV policy fixed; CAV-only GRPO does the converse. Full CoopLight updates both policies through joint GRPO.

Table 1. Effect of multi-agent co-training, averaged over jaywalking, incident, and barrier scenarios at both intersections (seconds).

MethodTLS updatesCAV updatesOverall AWT ↓Overall ATT ↓CAV AWT ↓CAV ATT ↓
SFT + communicationSFTSFT19.9046.0510.1835.10
TLS-only GRPOGRPOSFT18.8645.129.9534.71
CAV-only GRPOSFTGRPO19.4745.639.3834.22
CoopLightGRPOGRPO17.9544.288.9733.88

Joint co-training achieves the best overall and CAV efficiency, reducing Overall AWT by 9.8% over SFT + communication and outperforming both single-agent GRPO variants. Since GRPO updates both reporting and action selection, we further compare policies receiving the same ground-truth event types and locations. Co-training still reduces Overall AWT from 18.12 s to 17.21 s, indicating benefits beyond improved event reporting alone. These results support jointly learned, complementary decision making without requiring explicit action-level coordination.

Response to Weakness 2 & Question 2

Thank you for raising this concern. CoopLight focuses on intersection-level TLS–CAV cooperation under long-tail events, where multi-view semantic understanding, lane-level localization, and timely cross-agent communication are central challenges. Individual-intersection evaluation is well established in prior traffic-control research [1–8], particularly for investigating visually grounded decision making and complex local traffic interactions. Our focus is therefore complementary to network-wide traffic coordination, allowing us to examine cooperative decision making under long-tail events in greater detail.

Regarding simulation fidelity, while closed-loop deployment in the real world is currently restricted by safety regulations, our evaluation uses TranSimHub, integrating SUMO-based microscopic traffic simulation with Blender-based roadside and onboard camera rendering. Both environments reconstruct real-world intersection layouts rather than abstract grid junctions. Hong Kong Yau Ma Tei (YMT) features left-hand traffic, narrow roads, and 8 movements, while California (CAL) features right-hand traffic, wider multi-lane approaches, and 12 movements. The CAL network is reconstructed from OpenStreetMap, with lane connectivity, turning movements, routes, and signal phases manually verified. Each intersection includes 40 routes across four long-tail scenario categories under varying traffic densities, with disjoint training and evaluation routes, additional co-occurring events, and adverse weather and illumination conditions.

To evaluate cross-intersection transferability, we train on one intersection and evaluate on the other without target fine-tuning. Despite differences in driving conventions, road geometry, and signal configurations, CoopLight shows stronger transferability than VLMLight, with a smaller mean relative AWT degradation (14.9% vs. 25.5%; Table 2). We further evaluate two unseen intersections from the InTAS Ingolstadt and TAPAS Cologne SUMO scenarios without target fine-tuning. CoopLight maintains lower Overall AWT (20.0 s) than VLMLight (32.7 s) and CoopLight w/o Comm (23.9 s), suggesting that the benefit of semantic cooperation persists across different layouts. While these results support transfer across the evaluated intersection layouts, network-wide coordination and real-world deployment remain beyond our current scope. Extending semantic communication across adjacent intersections and evaluating network-level traffic propagation remain important directions for future work.

References

[1] Wang et al., VLMLight: Safety-Critical Traffic Signal Control via Vision-Language Meta-Control and Dual-Branch Reasoning Architecture, NeurIPS 2025.

[2] Yao et al., Incorporating Vision-Based Artificial Intelligence and Large Language Model for Smart Traffic Light Control, Applied Soft Computing, 2025.

[3] Afshari et al., Safety-Savvy Intersection Control Harnessing Vision Language Model, VEHITS 2026.

[4] Jeon et al., Artificial Intelligence for Traffic Signal Control Based Solely on Video Images, Journal of Intelligent Transportation Systems, 2018.

[5] Chu et al., Traffic Signal Control Using End-to-End Off-Policy Deep Reinforcement Learning, IEEE T-ITS, 2022.

[6] Liu et al., Integrated Optimization of Traffic Signal Timings and Vehicle Trajectories Considering Mandatory Lane-Changing at Isolated Intersections, Transportation Research Part C, 2024.

[7] Sun et al., Joint Optimization of Signal Control and Speed Planning at Isolated Mixed-Traffic Signalized Intersections, Applied Soft Computing, 2025.

[8] Lin et al., LLM-Driven Signal Control Method for Signalized Intersections with Mixed Traffic Flow, Systems, 2026.

Response to Question 3

Thank you for raising this question. We evaluate safety alongside traffic efficiency under identical traffic demand and event schedules across jaywalking, incident, and barrier scenarios at both intersections. To assess maneuver safety, we evaluate vehicle–vehicle and vehicle–pedestrian collisions, time-to-collision (TTC) conflicts with a minimum TTC below 3 s, and CAV trip completion rates, with SUMO safety checks disabled. Collision and TTC-conflict frequencies are normalized per 100 executed CAV lane changes.

Table 1. Safety evaluation with safety checks disabled, averaged across three long-tail scenarios at both intersections. Collisions and TTC conflicts are reported per 100 executed CAV lane changes.

MethodOverall AWT / ATT ↓CAV AWT / ATT ↓Vehicle–vehicle collisions ↓Vehicle–pedestrian collisions ↓TTC conflicts ↓Trip completion (%) ↑
CoTV27.36 / 46.2219.85 / 39.472.500.0018.286.7
CoopLight w/ SFT CAV w/o Comm22.71 / 48.1512.69 / 37.101.670.3316.193.3
CoopLight w/ SFT CAV w/ Comm20.14 / 46.3010.47 / 35.321.250.0015.695.6
CoopLight w/o Comm21.83 / 47.5211.84 / 36.211.170.1715.895.6
CoopLight18.21 / 44.499.26 / 34.070.700.0015.297.8

As shown in Table 1, CoopLight achieves 72.0% fewer vehicle–vehicle collisions and 16.5% fewer TTC conflicts than CoTV, while improving trip completion from 86.7% to 97.8%. Within the CoopLight variants, communication and co-training further reduce collision and conflict frequencies, suggesting that shared semantic information supports safer maneuver selection. These results indicate that CoopLight's traffic-efficiency gains are accompanied by improved safety outcomes in the evaluated closed-loop scenarios.

CoopLight produces high-level decisions, while SUMO handles vehicle motion. Lane-change requests are submitted through TraCI's changeLane interface and processed by SUMO's LC2013 model, which checks safe gaps to the target-lane leader and follower and allows cooperative speed adjustments. Infeasible requests are not executed. Longitudinal speed control and braking follow SUMO's Krauss car-following model, subject to speed limits and traffic signals. We further evaluate maneuver feasibility under SUMO's execution interface by reporting the proportion of blocked lane-change attempts alongside traffic efficiency.

Table 2. Intervention statistics under the SUMO execution interface, averaged across the same scenarios (seconds for AWT/ATT).

MethodOverall AWT / ATT ↓CAV AWT / ATT ↓Blocked lane changes (%) ↓
CoTV28.32 / 51.9515.17 / 39.8738.9
CoopLight w/ SFT CAV w/o Comm22.45 / 47.9612.38 / 36.8431.6
CoopLight w/ SFT CAV w/ Comm19.90 / 46.0510.18 / 35.1024.8
CoopLight w/o Comm21.57 / 47.2811.50 / 35.9727.3
CoopLight17.95 / 44.288.97 / 33.8818.5

Communication reduces the blocked-request rate for SFT CAVs from 31.6% to 24.8%, and full CoopLight lowers it to 18.5%. Relative to the TLS-only CoopLight variant, full CoopLight reduces Overall AWT/ATT by 25.8% / 11.9% and CAV AWT/ATT by 56.0% / 29.4%. Thus, the reported efficiency gains coexist with fewer rejected maneuvers and more favorable simulated safety outcomes. These improvements suggest that shared semantic information helps CAVs anticipate affected lanes and select feasible maneuvers before encountering heavily constrained traffic conditions. We will incorporate these safety evaluations and intervention statistics into the revised manuscript.

Thank you for your positive feedback. We will incorporate the additional experiments and clarifications into the revision. Thank you again for your time!

Comments on the paper

7 comments
AuthorsOpenReview ↗

Response to Multi-Intersection Scalability

Thank you for the follow-up. To address this concern, we evaluate CoopLight on the connected Cologne3 corridor from RESCO, which contains three consecutive signalized intersections: A (four-leg, 8 incoming lanes, 4 phases), B (four-leg with two single-lane side streets, 6 incoming lanes, 4 phases), and C (T-junction, 5 incoming lanes, 3 phases). The intersections are spaced 353 m (A–B) and 303 m (B–C) apart, with traffic demand drawn from the 07:00–08:00 morning peak of TAPAS Cologne. We construct roadside and onboard views using the same SUMO–Blender pipeline as in our original experiments. Each intersection runs its own TLS agent with local CAV cooperation; all agents operate simultaneously without target-network fine-tuning or additional TLS-to-TLS communication.

Vehicles traverse multiple intersections in five settings: an incident at intersection A (S1a), a barrier at intersection B (S1b), jaywalking at intersection C (S1c), all three events occurring together (S2), and an emergency vehicle passing through A–B–C (S3). We compare Max-Pressure, CoTV, VLMLight, CoopLight w/o Comm, and CoopLight. We report corridor-wide traffic efficiency and the duration of queue spillback between adjacent intersections.

Table 1. Overall AWT / ATT (s) on the connected Cologne3 corridor.

MethodS1a: Incident @AS1b: Barrier @BS1c: Jaywalking @CS2: Co-occurringS3: Emergency
Max-Pressure31.4 / 96.829.7 / 93.526.2 / 88.944.8 / 118.624.5 / 86.2
CoTV29.6 / 94.128.3 / 91.725.4 / 87.641.9 / 114.323.8 / 85.1
VLMLight33.5 / 99.431.8 / 96.227.9 / 91.347.6 / 122.925.9 / 88.0
CoopLight w/o Comm26.1 / 89.725.0 / 88.223.1 / 84.936.4 / 106.521.7 / 82.6
CoopLight22.3 / 85.221.9 / 84.621.2 / 82.729.8 / 98.120.4 / 80.9

Table 2. CAV AWT / ATT (s) and spillback duration (s). Spillback is the duration for which an inter-intersection queue reaches and blocks its upstream junction.

MethodS1a CAVS1b CAVS1c CAVS2 CAVSpillback S1aSpillback S1bSpillback S1cSpillback S2
Max-Pressure27.8 / 92.425.9 / 89.622.4 / 84.141.2 / 114.71189664241
CoTV23.1 / 87.522.6 / 86.421.0 / 83.035.7 / 107.21018859207
VLMLight30.2 / 95.628.7 / 92.824.9 / 87.744.9 / 119.612710971268
CoopLight w/o Comm19.8 / 83.619.1 / 82.418.7 / 80.929.4 / 99.2736247152
CoopLight14.6 / 78.114.9 / 78.316.3 / 78.521.7 / 90.631283579

Table 3. Emergency-vehicle traversal through A–B–C (S3).

MethodTraversal time (s) ↓Stops ↓
Max-Pressure78.42.0
CoTV74.91.7
VLMLight61.21.0
CoopLight w/o Comm55.70.7
CoopLight49.30.3

Across single-event scenarios (S1a–S1c), CoopLight consistently reduces Overall AWT and spillback compared with CoTV, with smaller gains for transient jaywalking events. By prioritizing dischargeable movements and enabling earlier CAV lane changes, CoopLight reduces spillback from 88 s to 28 s in S1b, preventing queues from extending into upstream intersections.

The advantage becomes more pronounced when events co-occur (S2), where CoopLight reduces Overall AWT by 28.9% relative to CoTV, with further benefits from semantic communication. CAV reports provide complementary evidence when local views are occluded, helping reduce downstream congestion. In the emergency scenario (S3), successive local priority decisions shorten traversal time to 49.3 s, compared with 61.2 s for VLMLight, with fewer stops.

These results show that local TLS–CAV cooperation remains effective under interconnected traffic dynamics, reducing spillback without explicit TLS-to-TLS coordination. Larger-scale networks and inter-intersection semantic communication remain future directions. We will include this corridor evaluation in the revision.

Thank you for your prompt feedback. However, three intersections do not meet the scalability requirement. If you have time, please use datasets from previous studies (e.g., Jinan, Hangzhou, NYC) [1]; they include at least 12 intersections for evaluation.

[1] Wei, Hua, Nan Xu, Huichu Zhang, Guanjie Zheng, Xinshi Zang, Chacha Chen, Weinan Zhang, Yanmin Zhu, Kai Xu, and Zhenhui Li. “CoLight: Learning Network-Level Cooperation for Traffic Signal Control.” In Proceedings of the 28th ACM International Conference on Information and Knowledge Management (CIKM ’19), 1913–1922. New York: Association for Computing Machinery, 2019.

Thank you for the suggestion. We clarify that CoopLight focuses on heterogeneous cooperation between infrastructure-side traffic signals and vehicle-side CAV agents under long-tail events, rather than explicit coordination among neighboring traffic signals. Such events, including incidents, barriers, and jaywalking pedestrians, typically involve localized disruptions that require timely scene understanding and complementary TLS–CAV decisions. To further evaluate scalability beyond the Cologne3 corridor, we extend our evaluation to larger connected networks.

Since CoopLight relies on roadside and onboard visual observations, the CoLight benchmarks [1] require additional visual scene construction beyond their original traffic-state representations. To address scalability, we evaluate CoopLight on two real-world-derived networks from CoLight: Jinan (3×4, 12 signalized intersections) and Hangzhou (4×4, 16 signalized intersections), using the Jinan-1 and Hangzhou-1 traffic demand profiles. We reconstruct both networks in SUMO and generate roadside and onboard views using the same SUMO–Blender pipeline as in our original experiments. Each intersection runs its own TLS agent with local CAV cooperation. All 12 or 16 TLS agents operate simultaneously without target-network fine-tuning or TLS-to-TLS communication.

Following the Cologne3 evaluation, we consider five settings: an incident (S1a), a barrier (S1b), and jaywalking (S1c), each introduced at a randomly selected intersection; three co-occurring events (S2); and an emergency vehicle traversing four consecutive intersections (S3). We compare Max-Pressure, CoTV, VLMLight, CoopLight w/o Comm, and CoopLight, reporting network-wide AWT/ATT, CAV efficiency, spillback, and emergency-vehicle traversal performance.

Table 1. Overall AWT / ATT (s) on Jinan (12 intersections).

MethodS1a: IncidentS1b: BarrierS1c: JaywalkingS2: Co-occurringS3: Emergency
Max-Pressure53.3 / 291.452.6 / 290.746.2 / 283.961.8 / 304.845.9 / 281.6
CoTV51.9 / 290.249.4 / 288.845.3 / 282.560.7 / 303.343.8 / 279.9
VLMLight54.8 / 295.153.9 / 292.048.9 / 286.265.1 / 309.647.6 / 284.0
CoopLight w/o Comm49.1 / 286.547.8 / 284.345.4 / 281.056.2 / 296.444.1 / 278.8
CoopLight45.0 / 279.446.1 / 281.543.7 / 280.151.9 / 290.242.5 / 277.9

Table 2. Overall AWT / ATT (s) on Hangzhou (16 intersections).

MethodS1a: IncidentS1b: BarrierS1c: JaywalkingS2: Co-occurringS3: Emergency
Max-Pressure56.7 / 306.855.6 / 304.750.6 / 298.464.9 / 318.849.7 / 298.1
CoTV54.5 / 304.852.8 / 301.950.1 / 297.262.2 / 313.647.6 / 294.9
VLMLight58.6 / 310.457.3 / 307.852.9 / 303.166.9 / 323.751.4 / 302.2
CoopLight w/o Comm52.6 / 301.351.3 / 299.649.2 / 296.858.8 / 309.948.0 / 294.7
CoopLight50.7 / 298.749.8 / 297.448.5 / 295.954.6 / 304.047.3 / 293.8

Table 3. CAV AWT / ATT (s) and network-level spillback duration (s) under co-occurring events (S2). Spillback is summed across internal links.

MethodJinan CAVJinan SpillbackHangzhou CAVHangzhou Spillback
Max-Pressure61.4 / 302.941263.9 / 316.5478
CoTV54.3 / 294.137158.6 / 309.6430
VLMLight64.2 / 306.544866.9 / 322.4521
CoopLight w/o Comm51.8 / 292.029654.2 / 303.3338
CoopLight44.2 / 281.618747.8 / 295.3221

Table 4. Emergency-vehicle traversal time (s) / average stops under S3.

MethodJinanHangzhou
Max-Pressure198.5 / 2.7262.4 / 3.0
CoTV192.1 / 2.3255.7 / 2.7
VLMLight171.6 / 1.3232.9 / 1.7
CoopLight w/o Comm163.0 / 1.0223.5 / 1.3
CoopLight151.8 / 0.3210.6 / 0.7

Across both networks, CoopLight achieves the lowest AWT and ATT in all five settings. Under co-occurring events (S2), it reduces Overall AWT by 14.5% on Jinan and 12.2% on Hangzhou relative to CoTV, with further gains from semantic communication and reduced network spillback. In S3, local emergency-priority decisions shorten traversal times to 151.8 s and 210.6 s, respectively, with fewer stops. These results demonstrate the effectiveness and scalability of CoopLight across connected networks with 12 and 16 intersections. We will include these evaluations in the revised manuscript.

[1] Wei et al. CoLight: Learning Network-Level Cooperation for Traffic Signal Control. CIKM 2019.

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

Related papers

Discussion (0)

Sign in to comment.