acceptodds
Under review as a conference paper at ICLR 2027

CARLAGen4D : A Multi-Agent Framework for Text-Driven 4D Scene Generation in CARLA

Abstract

Autonomous driving research heavily relies on high-fidelity simulators such as CARLA, where evaluation is almost exclusively performed on a limited set of predefined, artist-designed towns. Prior work in simulation-based evaluation largely depends on these fixed maps, constraining scenario diversity and limiting the ability to test generalization to novel environments and edge cases. As a result, current simulation-based evaluation pipelines struggle to scale with the complexity and variability required for robust autonomous driving validation. To address this gap, we present an end-to-end framework for text-to-4D simulation generation that directly enables controllable, scenario-driven evaluation of AD systems. Given a natural language prompt describing a testing scenario (e.g., unusual road topology, adversarial traffic behavior), our multi-agent system generates a complete, temporally consistent simulation, jointly synthesizing road topology, background assets, and dynamic agent behaviors from scratch. This removes reliance on fixed maps and enables on-demand creation of targeted evaluation scenarios. Since direct map generation is prohibitively expensive in token space, we design a domain-specific language that enables efficient map synthesis. We further benchmark driving policies across diverse generated road layouts and traffic scenarios, revealing their sensitivity to map configuration and scenario diversity. Beyond evaluation, the generated scenes support a range of downstream applications, including RGB-to-BEV perception, scene narration and captioning, and transfer to real-world imagery.

Primary area:
applications to robotics, autonomy, planning
Keywords:
4D scene generation, autonomous driving, CARLA, multi-agent LLM systems

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_so3YOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper introduces CARLAGEN4D, a framework that converts text prompts into executable CARLA simulation environments by generating road layouts, background assets, and actor trajectories. A compact DSL and compiler feedback support map construction and repair. Authors provide experimental results on a benchmark for text-to-map generation, covering over 250 prompts, and test 7 driving policies on held-out generated maps to demonstrate the downstream application of the proposed method. Additionally, the paper shows that the generated environments could be used as training data for both driving policy training, perception model training and reasoning model training, showing the potential of the proposed framework.

Critical strengths
  1. The paper is focusing on the useful topic for testing / evaluation for autonomous driving. Generating the road network itself addresses a practical limitation of systems that place scenarios in existing towns. And the experimental results are encouraging, CARLAGEN4D produces valid map artifacts for 98.4% of the 250 prompts and receives a VLM score of 4.54, compared with 4.02 for OmniTester. This capability makes it possible to construct environments around a requested interaction rather than search for a suitable location in a fixed map.
  2. The ablation studies provide clear evidence for the DSL and compiler feedback. Under a common model and evaluation budget, the complete method produces 39 clean maps out of 50, compared with 12 when generating raw OpenDRIVE and 33 without iterative repair. The DSL also reduces output tokens by 2.6 times and generation time by 2.1 times relative to direct OpenDRIVE generation, demonstrating both the effectiveness and efficiency of the proposed framework.
  3. The downstream training experiments across different tasks help demonstrate the practical value of the generated scenes. The authors evaluate waypoint prediction, camera to BEV perception, and driving narration and justification. With a matched training frame budget, generated data reduces waypoint error from 7.79 to 3.10 m and improves BEV mIoU from 0.110 to 0.364 on unseen generated maps, while also improving performance on existing CARLA towns. These results suggest that the framework provides useful training coverage beyond its role in scenario generation.
Critical weaknesses
  1. The distinction between producing a valid map and fully satisfying the prompt needs to be clearer. The main evaluation reports 98.4% valid artifacts, while the separate ablation reports clean completion on 39 of 50 prompts. These results use different criteria and prompt distributions, so they are not directly contradictory. However, returning an earlier map after budget exhaustion may leave requested features incomplete. Similarly, the 69.4% elevation score measures whether any grade is present, rather than whether the requested elevation structure is correct.
  2. The trajectory comparison largely reflects whether a suitable map is available. CARLAGEN4D executes all 39 scenarios, whereas TTSG executes 12, but 26 of TTSG’s 27 failures occur before simulation begins. This supports the value of generating compatible road networks, but provides less evidence that the trajectories themselves are better. Motion fidelity scores are close, at 3.06 and 2.99, with overlapping confidence intervals.
  3. The policy benchmark does not fully isolate why some maps are harder. Across 15 generated maps, InterFuser’s driving score decreases from 0.410 on easy layouts to 0.123 on elevated layouts. This is a useful observation, but differences in traffic, route length, appearance, and geometry could also contribute. More controlled comparisons would strengthen the interpretation that elevation is the main source of failure.
  4. The downstream task is not fully evaluated, in particular, the driving policy training results are evaluated only through open loop prediction. Reducing waypoint error from 7.79 to 3.10 m is encouraging, but it does not establish that the trained policy completes more routes or avoids more collisions. The separate evaluation of 7 released policies demonstrates the difficulty of the generated maps rather than the driving benefits of training on them.
  5. Minor concern: the photorealistic video examples do not yet demonstrate transfer to real driving data. No quantitative real data evaluation accompanies the translated videos, and Figure 17 shows clear artifacts like pedestrian's appearance is totally wrong. This raises a practical question about whether simulator annotations remain accurate after translation, particularly for dynamic objects.
Questions for authors
  1. Could the authors clarify how many of the 250 prompts satisfy all requested constraints and route checks? In particular, how are maps returned after budget exhaustion counted in the 98.4% generation success rate?
  2. Do the authors have existing closed loop results for policies trained with generated data? These would clarify whether the waypoint error reduction from 7.79 to 3.10 m translates into better route completion or fewer collisions.
  3. Could the authors provide trajectory quality results on the 12 scenarios that both methods successfully execute? This would help separate improvements in trajectory generation from the benefit of having a compatible map.
  4. Could the authors explain how route length, traffic, and other difficulty factors vary across the 15 evaluation maps? This would clarify whether the observed performance drops can be attributed specifically to road topology.
10 comments

Q1/W1 - "Valid map" vs. "satisfies the prompt"; how budget exhausted maps are counted - We agree that a loadable map does not necessarily satisfy all requested constraints. We therefore provide a breakdown of the 250 prompts using progressively stricter criteria. We now report the number of maps that satisfy different constraints and route checks. The reported 98.4% generation success rate measures whether the generated artifact can be loaded by CARLA, not whether the agent fully satisfies the prompt. In particular, 12 maps were returned after budget exhaustion; 7 of these have prompt-adherence scores <= 3, highlighting that artifact validity alone does not guarantee prompt satisfaction. Of the remaining 234 maps for which generation completed cleanly, 210 pass route checks and contain a single connected lane network. Further, 183 achieve both prompt-adherence and, for elevation-related prompts, elevation scores >= 4. We have added this breakdown in Appendix F.3 of the revised paper.

Criterion (each row adds to the one above)prompts%
Loadable map (reported 98.4%)24698.4
Final compile: 0 errors (agent finished cleanly)23493.6
+ single connected lane network21084.0
+ prompt adherence >= 4 and, for vertical prompts, elevation >= 418373.2

Q2/W4 - Closed-loop policy evaluation results - We additionally evaluate the driving-policy checkpoints from Table 3 in closed-loop simulation on CARLA test towns. Adding generated data improves both route completion and collision rates. These results provide evidence that the open-loop waypoint improvements translate to downstream closed-loop driving benefits.

We emphasize that these are small, single-camera policies trained offline at the scale used in Table 3. Their absolute route completion remains low, and we do not claim that generated data alone produces highly competent driving policies. Rather, these results establish that generated data can improve downstream policy performance under our matched data experimental setup. We also report these closed loop metrics in Appendix H.4.2 of the revised paper.

Training dataRoute completion ↑Collisions / km ↓
Hand-authored only0.013 ± 0.00330.3 ± 3.0
+ generated (matched size)0.116 ± 0.02022.9 ± 3.1
+ generated (all data)0.127 ± 0.02619.4 ± 0.7
Generated only0.07223.5

Q3/W2 - Trajectory quality on the 12 scenarios both methods execute - We agree that the original comparison is influenced by map compatibility. To isolate trajectory quality, we additionally evaluate both methods on the 12 scenarios that are successfully executed by both. On this common subset, CARLAGen4D achieves higher motion fidelity and lower heading discontinuity but this is within overlapping confidence intervals. Therefore, even though ours is slightly better, we do not claim strictly better trajectory quality when both methods run. The advantage of our method is executability i.e. it runs all 39 prompts against TTSG's 12, because it builds a road network for each prompt instead of searching a stock town for a matching site. Please also find the analysis in Appendix G.3. of the revised paper.

MetricOursTTSG
Motion fidelity3.212.99
Collision rate0.0000.071
Heading discontinuity (%)0.131.266
Ego-actor collision0.0000.077
Reviewer_r9KyOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CARLAGen4D converts natural-language descriptions into executable driving scenes in CARLA. A road-topology agent writes programs in a compact Python-embedded DSL and iteratively repairs them using compiler diagnostics and route checks. Background agents place static assets, while foreground agents place traffic participants and generate motion specifications that are checked through simulation. The paper evaluates map generation against four baselines on 250 prompts, benchmarks seven driving policies on 15 generated maps, and studies generated data for waypoint prediction, camera-to-BEV perception, and driving captioning. The results suggest that the DSL improves generation efficiency and that generated data can complement hand-authored CARLA towns.

Critical strengths
  1. A concrete, executable system for building complex driving environments. Rather than merely producing textual scene descriptions, the work implements road representations that explicitly encode connectivity and elevation, combined with route validation relying on both compiled graph structures and CARLA’s native parser. Additional modules handle actor placement and rollout verification. The ablation study in Appendix F.4 demonstrates the practical merit of this implementation: the full pipeline substantially outperforms simpler variants, especially compared to directly generating OpenDRIVE XML under identical computational constraints. The system can support targeted testing for rare and nonstandard road layouts.

  2. The paper presents evidence that synthetic scenes enrich training coverage. In Table 3, half of the manually crafted training data is substituted with generated scenes while keeping the total training frame count fixed. The model sees clear gains in waypoint prediction on unseen generated maps, alongside mild improvements on hand-built towns; camera-to-BEV performance also improves accordingly. By holding out full maps and evaluating across multiple random seeds, the authors rule out trivial explanations such as extra training samples or geometry leakage between train and test splits. While the waypoint evaluation is open-loop, these simulation findings remain valuable.

Critical weaknesses
  1. The contribution of the proposed DSL over existing high-level road-generation APIs remains unclear. The ablation against direct OpenDRIVE XML generation demonstrates its practical value, but does not isolate the benefit of this particular DSL design. Since prior methods also use road-generation tools, the paper should clarify what the DSL adds in expressivity or generation reliability.

  2. The external comparisons lack sufficient control and reporting of generation costs. The internal ablations use the same model and time budget, but OmniTester and ML-SceGen are run with their default configurations. Their token consumption, LLM call counts, and repair iterations are not reported. It is therefore unclear whether the performance advantage holds under comparable inference budgets.

There is also an error in the extent comparison on page 7, lines 351–352. The text states that OmniTester trails on extent, although its value of 1.20 is closer to the stated target of 1 than the proposed method’s 1.28.

Questions for authors
  1. What advantages does the proposed DSL offer over existing high-level road-generation APIs? Could the authors replace the DSL with an existing high-level API, rather than direct OpenDRIVE XML generation, and compare the results with all other conditions unchanged?

  2. Which models and generation budgets were used for the proposed method, OmniTester, and ML-SceGen? Please report the input/output token consumption, LLM call counts, repair iterations, and runtime associated with the reported results. Could generation quality also be compared under matched budgets on a representative subset?

3 comments

Q1,W1 - What the DSL adds over an existing high-level road API - We agree that the comparison against direct OpenDRIVE XML generation alone does not isolate the benefits of our DSL design. To address this, we conduct an additional experiment replacing only our DSL with the SUMO netconvert node/edge API, the same high-level API used by OmniTester, while keeping the agent, underlying model, compile-and-repair loop, feedback mechanism, and time budget unchanged. In particular, we give the agent the full API specification with a worked example, and feed compilation errors back after each iteration just as in our original pipeline. The results on a representative subset of 30 prompts are provided in the table below.

Representation (same agent and loop)Maps passing all structural checksSelf-intersecting mapsPrompt adherenceElevation (vertical prompts)Output tok / promptInput tok / prompt ( % cache reads)
Our DSL27/3033.804.3356.5k176.5k (85%)
SUMO high-level API17/30133.531.6738.6k197.5k (92%)
(OmniTester, for reference)16/30143.701.4359.6k33.0k (12%)

Our DSL achieves 27/30 structurally valid maps, compared to 17/30 with SUMO, and substantially better elevation generation (4.33 vs. 1.67). This difference stems from how the two representations encode road geometry and connectivity. In SUMO's node/edge representation, the agent specifies junction points and the road segments connecting them, leaving the compiler to infer their geometric connections. This can silently produce overlapping or self-intersecting roads. In contrast, our DSL explicitly specifies road-to-road connectivity and constructs geometry using OpenDRIVE primitives (lines, arcs, and spirals). These additional results are now reported in Appendix F.6 of the revised paper.

Q2,W2 - Models, budgets and cost, and quality at matched budget - All three methods (our proposed method, OmniTester, and ML-SceGen) use the same LLM, Claude Sonnet 5 as their backend. We additionally report generation costs on the same 30-prompt subset, including token consumption, number of LLM calls, repair iterations, and runtime.

MethodOutput tokInput tok (of which cache reads)LLM callsRepair iters, meanRuntime medianCost / prompt
Ours56.5k176.5k (85%)8.11.9367 s\$0.66
OmniTester59.6k33.0k (12%)6.60.7347 s\$0.67
ML-SceGen2.3k2.1k3.2020 s\$0.03

We enforce a time budget of 1200s for all methods here. Although our method consumes approximately 5x more raw input tokens than OmniTester because it re-sends the growing conversation at each call, 85% of these tokens are served from the prompt cache. Therefore, using Claude Sonnet 5 pricing (\2/Minput,2/M input, \\2.50/M cache writes, \0.20/Mcachereads,and0.20/M cache reads, and \\10/M output tokens), the effective cost is comparable: \0.66vs.0.66 vs. \\0.67 per prompt. Output tokens account for approximately 85-89% of the total cost for both methods. To further address the budget concern, we compare generation quality under a matched and reduced budget of 51k output tokens per prompt. The cost analysis is reported in Appendix F.6 of the revised paper.

Extent ratio clarification - We thank the reviewer for catching this error. The main-text statement that OmniTester trails in extent is incorrect. OmniTester's ratio (1.20) is indeed closer to the ideal value of 1 than ours (1.28). We have clarified this in Section 4.2 of the revised paper.

Generation Quality at Matched and Reduced Budget - Under the matched output budget, we reduce our budget to 51k tokens (median token count of OmniTester) our method retains its advantage in structural validity (24 vs. 16) and elevation generation (4.12 vs. 1.43), while achieving comparable prompt adherence (3.53 vs. 3.70). The reduction in our structural validity from 27/30 to 24/30 under the cap is due to prompts whose generation was truncated by the budget limit. These results show that our performance advantage over OmniTester persists under comparable output-token and monetary budgets, rather than arising simply from higher inference expenditure. These additional results are now reported in Appendix F.6 of the revised paper.

Method (on 30 prompts)Output tok / promptCost / promptMaps passing all structural checksPrompt adherenceElevation (vertical prompts)
Ours, uncapped56.5k\$0.6627/303.804.33
Ours, capped at 51k47.2k\$0.5524/303.534.12
OmniTester59.6k\$0.6716/303.701.43
ML-SceGen, its default2.3k\$0.0324/300.570.00
AuthorsOpenReview ↗

Followup on our response

Dear Reviewer r9Ky,

Thank you again for your review. As the discussion period is closing soon, we wanted to check whether our responses have addressed your concerns. In brief,

  1. W1,Q1 - Swapping only our DSL with an existing high-level road API i.e. SUMO netconvert API gives 17/30 valid maps versus 27/30 with our DSL.
  2. W2,Q2 - All methods use the same LLM. We get a similar cost with respect to the closest baseline ($0.66 ours vs. $0.67 OmniTester per prompt), and under a matched output budget we still lead OmniTester 24/30 to 16/30.
  3. W2 - Extent comparison - corrected in Section 4.2.

If any point remains unclear, we would be glad to clarify it in the remaining time.

Warm regards, The Authors

Reviewer_XuwCOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents CARLAGen4D, an LLM-agent pipeline that turns a text prompt into a CARLA scene. A road-topology agent writes a program in a Python-embedded DSL that is compiled to OpenDRIVE, with compiler diagnostics and route checks fed back for repair. Separate agents then place background assets on non-drivable plots and generate actor trajectories that are validated by rollout. Map generation is compared against four text-to-map baselines on 250 prompts using mechanical metrics and a VLM judge. The generated maps are further used to benchmark seven driving policies on 15 held-out maps, and as additional training data for an open-loop driving policy, a camera-to-BEV model, and the ADAPT captioner.

Critical strengths
  1. Generating the map itself, rather than placing actors on a stock town, is the right step beyond ChatScene- or TTSG-style systems, and the paper shows the potential. All generated scenarios could execute, against 30.8% for TTSG, whose failures mostly come from the fixed town lacking a suitable site (Sec. 3.2). The DSL is a good design choice compared with generating OpenDRIVE directly, which gives higher VLM score with fewer output tokens.
  2. The evaluation includes controls that many papers in this area skip. The shuffled-prompt gap in Sec. 4.2 and the matched frame budget in Sec. 4.4 both rule out easy explanations for the gains. The shuffled-clip result in Table 4 is particularly telling: it shows the released ADAPT checkpoint barely looks at the video.
Critical weaknesses
  1. The policy benchmark (Sec. 4.3) is too thin for the claims made about it. It covers only 15 maps with no variance reported, and the category trends are not consistent: UniAD and VAD actually score higher on Junction than on Easy, and Roach scores highest on Elevation. More importantly, it reports no results for the same policies on stock CARLA towns under the same script. Without that reference, it is unclear whether the drops reflect topology or the visual shift introduced by the generated background. Running the script on one or two stock towns would help separate the two.

  2. The largest training-data gains in Table 3 are on held-out generated scenes, which are in-distribution for the augmented models. On hand-authored towns the gains are modest (L2 4.90→4.42, mIoU 0.157→0.183), and they come on top of weak baselines with 4–5 m L2. Also a minor error in Sec. 4.2: the text says OmniTester trails on extent, but its ratio (1.20) is closer to 1 than the proposed method's (1.28).

Questions for authors
  1. How many routes are there per map and per category in Sec. 4.3? How do the same policies score on stock CARLA towns under your script? Is foreground traffic present in these runs?
  2. The paper specifies Town02/05 as the validation split. Please clarify whether Table 3 reports results on this split or on a separate test set. A small-scale evaluation with a stronger driving policy like TF++, or using closed-loop driving metrics, would also help establish the broader value of the generated training data.
  3. Since generation currently relies on claude-sonnet-5 at about 10 minutes per map, have you tried an open-weight model, or fine-tuning a smaller one on the verified DSL programs?
6 comments

W1&Q1 - Policy benchmark, stock towns, variance, category trends, routes and traffic -

Category trends, variance and route length - We thank the reviewer for pointing out these limitations. We have reported the variance across 3 seeds for all episodes and agents in the revised paper (in Table 2). The overall trend remains the same as before. The inconsistent category trends arise partly because route lengths are not matched i.e. junction routes are 2.3x shorter than easy routes (206 m vs 482 m on average). To control for this, we rescored all episodes over the first 200 m, keeping the evaluation protocol and scoring unchanged. The results are:

7-agent mean DSeasyjunctionelevation
Full route (Table 2)0.480.510.46
First 200 m0.720.490.57

At matched route lengths, junction routes are consistently harder than easy routes across all seven agents. In particular, the UniAD and VAD inconsistencies disappear (UniAD: 0.52 vs 0.47; VAD: 0.44 vs 0.33 for easy vs. junction). Roach's stronger performance on elevation is also expected, as its BEV-based representation is largely appearance-independent and its privileged inputs make it less sensitive to visual and slope-related challenges. We have added these new results and analysis in Appendix H.3.1 of the revised paper.

Routes and foreground traffic - There is one route per map, giving 15 routes in total, with five maps/routes per category. Every policy is evaluated on the identical routes. Only scenario actors are present in these evaluations. We have clarified this in the Protocol paragraph of Section 4.3 in the revised paper.

Comparison with stock CARLA towns - To isolate the effect of generated appearance, we additionally selected the five easy scenarios, for which we could identify the closest-matching road layouts in stock CARLA towns. We reconstructed these scenarios while retaining the same foreground scenario actors. Please note that comparable matches were not available for the other categories i.e. junction-heavy and elevation scenarios, whose more complex topologies have no exact or similar counterparts in the hand-authored towns. The results are in the table below for a subset of non-privileged agents including comparison with generated towns which we rerun.

AgentGenerated townsCARLA Hand-authored TownsChange
TF++0.750.85+0.10
InterFuser0.460.34-0.12
CILRS0.630.75+0.12
LBC0.330.27-0.06

The results show that generated appearance does not consistently handicap camera-based policies: TF++ and CILRS perform better on CARLA towns, while InterFuser and LBC perform better on generated towns. This suggests that the observed performance differences cannot be attributed solely to visual shift and that scenario and topology complexity also play an important role. We also perform an ablation without traffic to see the effect of only topology where we find that junction and elevation maps remain challenging even without traffic. Please find more details in Appendix Section H.3.1 of the revised paper.

W2 - Training-data gains are mostly in-distribution - The mIoU improvement understates the gains on these towns because it averages over 11 classes, five of which (stop sign, three traffic-light states, and walker) occupy only a few pixels at 0.5 m/px. All models score close to zero on these classes, limiting their contribution to the overall metric. On the remaining classes, adding generated data improves performance by 10-22% (Appendix, Table 20).

Our objective in this experiment is to isolate the contribution of generated training data, rather than introduce a stronger model or training recipe. Comparing hand-authored-only training against a matched-frame-budget mixture with generated data allows us to measure this contribution directly.

Closed loop metrics - We now report closed loop metrics for the policies trained in Section 4.4.1 in Appendix H.4.2 of the revised paper. We observe a similar trend in the route completion and number of collisions as seen in open loop metrics of Table 3. Note that existing strong policies such as TF++ and InterFuser have already been trained on large hand-authored CARLA datasets, making a comparable controlled comparison infeasible. We also clarify that these are small, single-camera policies trained offline and we do not claim that their results establish closed-loop competence.

Extent ratio clarification - We thank the reviewer for catching this error. The main-text statement that OmniTester trails in extent is incorrect. OmniTester's ratio (1.20) is indeed closer to the ideal value of 1 than ours (1.28). We have clarified this in Section 4.2 of the revised paper.

Q2 - Table 3 split - We thank the reviewer for pointing out the need to clarify the evaluation split which we will now make explicit in the revised paper (please see Table 3). Table 3 reports results on a separate held-out test set comprising Town06, Town07, and Town10HD, while Town02 and Town05 are used exclusively for validation, as specified in the paper. To make this distinction clearer, we additionally report validation results and per-town test results below. Importantly, the gains from incorporating generated data are consistent across both validation and held-out test sets. The improvement is also consistent across all three individual test towns which supports our central finding that generated scenes can complement hand-authored training data and improve generalization beyond the generated training distribution.

Training dataVal mIoUTest mIoUTown06Town07Town10HD
Hand-authored only0.1500.1570.1990.1210.151
+ generated, matched budget0.1890.1830.2360.1410.172
+ generated, all data0.1820.1720.2120.1270.177
Generated only0.0520.0820.1030.1080.035

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary and general response

We thank the reviewers for their constructive feedback. We are encouraged that reviewers recognized several strengths of our approach i.e. generating road networks rather than relying on fixed towns enabling scenarios tailored to specific interactions and rare layouts [XuwC, r9Ky, so3Y]; the DSL and compiler feedback improve generation efficiency and quality, outperforming direct OpenDRIVE generation under the same budget with fewer output tokens and less time [XuwC, r9Ky, so3Y]; generated scenes improve training coverage, leading to better waypoint prediction and camera-to-BEV perception at a matched frame budget [r9Ky, so3Y].

We ran the following new experiments (details in individual responses) to address the questions asked by the reviewers -

  1. Cost table for all methods i.e. input and output tokens, LLM calls, repair iterations and runtime. (r9Ky) - We report the cost for all methods and show that the per-prompt costs for CARLAGen4D and the closest baseline, OmniTester, are comparable with CARLAGen4D producing higher fidelity maps.
  2. Matched-budget comparison i.e. our method clamped to OmniTester's measured output budget. (r9Ky) - Under a fixed and reduced 51k output-token budget, CARLAGen4D produces 24 structurally valid maps versus 16 for OmniTester, retaining its advantage without additional token expenditure.
  3. Our DSL replaced by an existing high-level road API, the SUMO node/edge API that OmniTester builds on, with other pipeline components unchanged. (r9Ky) - Replacing our DSL with OmniTester's SUMO node/edge API while keeping the rest of the pipeline unchanged yields 17/30 valid maps, versus 27/30 with our DSL. Our advantages persist at nearly identical measured cost and under a stricter OmniTester output-token cap.
  4. Stock-town (Hand-authored CARLA towns) experiments, the same generated scenarios rebuilt on CARLA towns and their evaluation (XuwC, so3Y) - Reconstructed a subset of scenarios on hand-authored CARLA towns with the same actors. Results show that generated appearance does not consistently disadvantage camera-based policies.
  5. Per-map difficulty factors (appearance, route length and traffic) separation for Sec 4.3 (so3Y). - Separately evaluated appearance, route length, and traffic. Junction and elevation maps remain challenging even after controlling for these factors and removing traffic.
  6. Closed-loop driving metrics of the Table 3 policy checkpoints (XuwC, so3Y) - Evaluated the Table 3 policy checkpoints in CARLA. Adding generated data improves route completion and reduces collisions, supporting the downstream benefits beyond open-loop waypoint prediction.
  7. Our pipeline with an open-weight LLM (XuwC) - Evaluated Qwen3.8-Flash on 30 prompts, achieving 25 structurally valid maps versus 27 with Claude Sonnet 5, retaining most of the generation and elevation benefits.

We have uploaded a revised paper which addresses these questions. We hope that our responses and the revisions to the manuscript adequately address the reviewers' concerns and provide a clearer understanding of our contributions.

Dear Reviewers,

Please take time to review the responses from the authors and engage with them for further questions and comments if any.

Time left is not much and please do shortly.

Thanks

AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.