- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
VLALight uses a vision-language model with a 0.5B language backbone to select traffic signal phases. The model takes stitched camera views and a textual description of the phase configuration, then predicts an action through learnable queries and an action head. It is trained on simulated demonstrations from MaxPressure and an emergency-priority labeling rule. On the pooled regular-flow evaluation, it achieves the lowest emergency waiting time among the compared controllers, but its overall waiting time is worse than that of several baselines. The paper also evaluates inference latency, unseen traffic flows, held-out intersections, and the effects of image stitching and minimum green time.
Critical strengths
-Overall well-written and easy to follow.
-The latency result is a clear practical advantage over the cascaded VLM controller. A 224 ms average forward pass leaves substantial room within the five-second decision interval.
-The paper reports the cost of emergency prioritization to other traffic rather than showing only the favorable emergency metrics. The examples involving simultaneous emergencies and missed vehicles are also helpful for understanding the model's limitations.
Critical weaknesses
-I would like to see a simpler emergency-aware baseline. VLMLight is a relevant comparison, but a detector followed by MaxPressure with emergency preemption is also a natural design. Appendix F dismisses such a baseline as artificial, which I do not find convincing. An oracle version using simulator emergency identity would be useful even before adding a detector.
-The trade-off is quite large: emergency waiting falls from 12.5 to 9.84 seconds relative to VLMLight, while overall waiting rises from 11.2 to 19.2 seconds. Some penalty is expected when prioritizing emergency vehicles, but the current experiments do not show whether this much is necessary. It could also reflect errors in routine control.
-The experiments do not establish which parts of the architecture are needed. Much of the model comes from VLA-Adapter, and stitching changes resolution and token count as well as layout. A vision-only policy and a simpler classifier trained on the same data would help determine whether the language input and query mechanism contribute meaningfully.
-The latency and generalization claims need tighter support. The cloud timing includes overhead excluded from VLALight's forward-pass measurement, and it is unclear how inference time affects the simulation. The reported variation also comes from four event sets with one checkpoint, while the separation between training and evaluation emergency configurations is not fully explained.
Questions for authors
-How does the method compare with MaxPressure plus emergency preemption, first with oracle emergency identity and then with a visual detector?
-How well does VLALight perform when there are no emergency vehicles, and can it achieve a better routine-versus-emergency trade-off by changing the priority setting?
-What happens with a matched vision-only model or a simpler action head?
-How are inference delays incorporated into the simulator, which emergency configurations are held out, and how stable are the results across training seeds?
2 comments
Reply to Reviewer CCcw (Part 2)
Comment:
Below we address the component ablation, then latency inside the simulator, held-out configurations and stability across seeds.
[W3, Q3] Which parts of the architecture are needed? Would a vision-only policy or a simpler classifier trained on the same data perform as well?
We freeze the backbone and train a small classifier on its last-layer features. The features are the action-query states of the last LLM layer, mean-pooled into 896 dimensions. The probe uses 1,500 frames of the test split, drawn from the six intersections and split 80/20 by stratified random sampling into 1,200 fitting and 300 held-out frames. It tests whether the phase decision is already carried by the visual representation.
Logistic regression reaches 84.00% and a two-layer MLP 86.33%, while always predicting the most frequent phase of the fitting half reaches 30.00%. The generative head is therefore not the source of the capability. The phase is largely linearly decodable from the representation, and a simpler head suffices. Distilling the policy into such a classifier is the most concrete route to a cheaper deployment. What kind of action head signal control actually needs, and how far a lighter head can be pushed before it stops reacting to emergencies, is a direction our future work will focus on.
[W4, Q4] How are inference delays incorporated into the simulator, which emergency configurations are held out, and how stable are the results across training seeds?
The simulator is configured to take one decision every five seconds. While a decision is being computed its clock is frozen, and time starts running again only when the returned decision is executed. The reported metrics therefore contain no explicit delay term. A VLALight decision takes 224 ms on one RTX 4090, which is 4.5% of the interval, so inserting the real delay would change the numbers very little.
The freeze is where the comparison stops being realistic for the larger controllers. A frozen clock lets a slow method hide its cost, because the intersection does not move while it thinks. In reality the intersection keeps moving. VLMLight needs 10.9 s with its 32B backbone and 4.2 s with its 3B backbone against a 5 s interval, so an unfrozen loop would either skip decisions or execute a phase chosen for an observation one or two intervals old. That is why we report latency for every method rather than assuming the freeze: it flatters any controller whose decision time exceeds the interval, and none of ours does.
Emergency configurations are held out in a weaker sense than flows or topologies. Training uses the emergency frames present in the rendered scenes, with a randomised phase-to-view mapping. Testing uses four fixed configurations of six emergency vehicles each, with departure times and routes fixed in advance. The policy therefore meets emergency vehicles at positions and timings it never trained on, while the vehicle types themselves stay fixed.
Stability across seeds follows from the sampling regime. The training set contains about 16k decision frames, 1,462 emergency frames plus routine frames from 20 flows per intersection. Training runs 5,000 steps with an effective batch size of 128, which draws about 640k samples, or about 40 traversals of the set. Every frame is therefore seen about 40 times, and sampling noise from a different seed is averaged out. We report mean and standard deviation over four emergency sets with a single checkpoint, and the variation in our tables comes from the emergency configuration rather than from retraining.
Reply to Reviewer CCcw (Part 1)
Comment:
Below we report the preemption sweep and the no-event comparison. The sweep treats the sensing distance as the only free parameter; the no-event table runs every method with the injection sets disabled, on the same six intersections, with three seeds.
[W1, Q1] How does the method compare with MaxPressure plus emergency preemption, first with oracle emergency identity and then with a visual detector?
The preemption rule has one free parameter, the sensing distance: how many metres before the stop line an emergency vehicle becomes visible to the controller. At zero the rule never triggers, so the row reduces to plain MaxPressure and serves as the no-preemption reference.
Each distance is averaged over six intersections and four injection sets on the increasing-demand flow, 24 runs per distance.
Emergency waiting falls from 13.07 s without preemption to 6.1 s at 50 m, and routine waiting rises from 11.56 s to 14.96 s over the same range. The rule buys emergency clearance with routine throughput.
VLALight reaches 10.33 s from four images, with no detector and no radius to tune. It beats MaxPressure without emergency awareness at 13.07 s. It trails MaxPressure with preemption at 10 m, which reaches 7.99 s. The rule reads emergency identity from the simulator state, so it is an upper bound rather than a deployable baseline.
[W2, Q2] The trade-off is large: emergency waiting falls from 12.5 to 9.84 seconds while overall waiting rises from 11.2 to 19.2 seconds. How does VLALight perform without emergency vehicles, and can a better trade-off be obtained by changing the priority setting?
The first group below contains methods that perceive an emergency vehicle. The second contains methods that cannot.
Emergency-aware methods
Non-emergency-aware methods
VLALight's routine waiting is 19.15 s. It is above the four rule-based controllers and IntelliLight, which lie between 12.10 and 14.56 s, and that gap is the green time reserved for emergency vehicles. It is below the remaining eight learned controllers.
VLALight's routine waiting is 19.15 s. It is above the four rule-based controllers and IntelliLight, which lie between 12.10 and 14.56 s. That gap is the green time it holds back for emergency vehicles, so it is the price of the emergency capability rather than a loss on routine traffic. Against the learned controllers, which are the relevant reference for a policy trained from data, VLALight is the lowest of all of them, from AttendLight at 21.07 s down to Advanced-MPLight at 31.39 s. It also has the smallest spread in that group, 5.8 s.
Priority is expressed inside the instruction, not as a coefficient. There is no separate weight that could be turned up or down. How this priority should be adjusted, and how to improve the trade-off itself, is a direction our future work will focus on.