acceptodds
Under review as a conference paper at ICLR 2027

CornerBench: Is Your Driving Model Safe Enough?

Abstract

Modern end-to-end driving planners have improved rapidly on common driving benchmarks, but whether they can safely handle rare, high-consequence corner cases remains unclear. Evaluating such cases is difficult because natural driving logs contain few safety-critical events, while staged data collection is costly and difficult to scale. We present **CornerFactory**, an automated pipeline for constructing corner cases from real-world driving logs according to structured scene specifications. CornerFactory supports multi-view-consistent object insertion, novel-view synthesis from perturbed ego poses, and transfer across vehicle platforms with different camera configurations. Building on this pipeline, we introduce **CornerBench**, a benchmark that organizes generated corner cases into a structured taxonomy and evaluates planners using continuous, severity-aware metrics that capture graded safety differences. Using CornerBench, we evaluate representative planners along several dimensions commonly expected to improve robustness, including training data volume and model size, multimodal sensing, language-based scene understanding and chain-of-thought reasoning in VLA planners, and the use of world models for future prediction and planning.

Primary area:
datasets and benchmarks
Keywords:
Autonomous Driving, End-to-End Planning, Corner Case Generation, Safety Evaluation, Synthetic Data

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_8mccOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

Modern end-to-end driving planners perform well on common driving benchmarks, but it remains difficult to evaluate their true safety under rare corner cases. This is primarily constrained by the scarcity of safety-critical events in natural driving logs, the high cost of staged data collection, and the difficulty of existing generative methods in guaranteeing multi-view and cross-modal (LiDAR) geometric consistency. To address this, the paper proposes CornerFactory, an automated, specification-driven pipeline grounded in real-world driving logs. Utilizing 3D Gaussian reconstruction and ray tracing, it achieves multi-view-consistent object insertion, novel-view synthesis from perturbed ego poses, and vehicle-configuration transfer, while strictly providing aligned LiDAR updates. Building on this, CornerBench, a benchmark comprising 230 cases (spanning 3 major categories and 19 concrete scenario types) is constructed. By introducing continuous, severity-aware metrics, it enables a systematic and fine-grained quantitative evaluation of existing models' true robustness and safety response capabilities in long-tail scenarios.

Critical strengths
  1. Rigorous Geometric and Cross-Modal Alignment: By leveraging 3D Gaussian reconstruction and ray tracing, it ensures spatial self-consistency across multi-view and RGB-LiDAR inputs. This effectively reduces confounding factors where models might fail simply due to mismatched input interfaces or generation artifacts.

  2. Discriminative Evaluation Metrics: The continuous metrics can quantitatively distinguish between "unmitigated collisions" and "attempts to avoid hazards (e.g., significant deceleration)", providing a more fine-grained basis for evaluating a model's safety response strategies, rather than treating all failures identically.

  3. Exposing Limitations of Existing Methods: It systematically validates the bottlenecks of current purely data-driven models in long-tail scenarios, providing clear testbeds and strong motivation for introducing new mechanisms, such as causal alignment or experiential memory.

Critical weaknesses

None

Questions for authors

Q1. Open Source and Extensibility: Will the CornerFactory generation pipeline and the CornerBench dataset be fully open-sourced? Furthermore, does the pipeline support users in defining new structured scene specifications (e.g., specific causal intervention logic) to extend and generate novel test cases?

Q2. Fine-Grained Data Export: Does the evaluation script support exporting time-series data of intermediate states ? This would facilitate a more detailed analysis of the model's specific response behaviors at critical moments, rather than relying solely on the final aggregated score.

2 comments
AuthorsOpenReview ↗

Response to Reviewer 8mcc

We sincerely appreciate your insightful questions and we believe they hold significant value for our work. We try to resolve your concerns below.

Q1: Open source and structured extensibility

Thank you for raising this important question about openness and extensibility. We have now publicly released anonymous resources for inspection and reuse:

(1) Defining new scenarios. Users can edit structured specifications or sample parameter ranges with recorded random seeds:

SpecificationExamples of supported controls
Inserted objectCategory, dimensions, position
Temporal behaviorTrajectory, motion direction, timing
Ego-state interventionPerturbed ego poses

These controls let users define new corner cases grounded in recorded driving scenes. For example, they can vary an obstacle's position or crossing time to create different hazard scenarios. This flexibility enables the community to build customized datasets for model training, evaluation, and other research applications.

(2) Pipeline upgrades and dataset expansion. To provide the community with higher-quality generation tools and a more comprehensive dataset, we have made two updates:

  • Pipeline upgrade: we upgraded both NeoObj and NeoView from 336p (560×336) to newly trained 720p (1280×720) models and refined the generation pipeline.
  • Dataset expansion: using the upgraded pipeline, we generated additional corner-case data and expanded the evaluation set. The expanded experimental set is summarized below:
Case categoryOriginal setExpanded experimental setScale increase
Total2301,3005.7×
Object / scene editing2131,2005.6×
Ego-state perturbation171005.9×

We also compare video quality between the original and upgraded pipelines separately for ego-state perturbations and object insertion, using VBench components [R1]:

MeasurementEgo-state: oldEgo-state: newObject insertion: oldObject insertion: new
Imaging Quality ↑0.59150.63130.55210.5979
Subject Consistency ↑0.89890.92190.92360.9259
Motion Smoothness ↑0.96640.97450.98090.9797

Q2: Export of intermediate time-series states

Thank you for this suggestion. Our evaluation tooling saves both the predicted trajectory and the simulated ego-vehicle state sequence for each sample, allowing researchers to examine how a planner responds at critical moments. The available outputs support the following analyses:

Available outputSupported analysis
Predicted and executed trajectoriesComparing planned motion with simulated execution
Speed traces derived from saved statesInspecting when and how much the vehicle slows down
Per-sample collision statisticsIdentifying collision timing, impact speed, and speed reduction
Camera and BEV visualizationsRelating the planner's responses to the surrounding scene

For example, two planners may both collide yet differ substantially in how much they slow before impact. Their speed traces and collision statistics reveal these differences and help explain the behavior underlying the final score. For a concrete example, Appendix I, Figure 25 (p. 27) of the revised manuscript visualizes the predicted trajectories and simulated speed curves of Drive-JEPA and DiffusionDrive in three collision cases.

We appreciate your constructive feedback and hope our responses have addressed your concerns. We welcome any additional questions or suggestions that could help us further improve our work.

References

[R1] Huang et al. VBench: Comprehensive Benchmark Suite for Video Generative Models. CVPR 2024, pp. 21807–21818. https://arxiv.org/abs/2311.17982

Thank you for the additional clarifications. I will maintain my rating.

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

The paper introduces CornerFactory, a pipeline that edits real driving logs to create rare objects, environmental hazards, and perturbed ego states. It combines front-view video editing, 3D reconstruction and cross-view completion, novel-view synthesis, vehicle-camera transfer, and synthesized LiDAR updates. CornerBench contains 230 generated cases covering 19 scenario types from 125 nuPlan test scenes, with hazard-response, lane-keeping, and traffic-rule scores that distinguish degrees of failure. A broad planner comparison finds low scores on many cases and illustrates mismatches between language descriptions and planned actions.

Critical strengths
  1. The construction pipeline addresses practical obstacles to evaluating heterogeneous planners on shared corner cases. Multi-camera insertion, ego-pose perturbation, camera-configuration transfer, and LiDAR updates make the benchmark more useful than single-view appearance editing alone in Sections 3.2–3.5. Appendix A explains how task-specific completion branches address artifacts of cross-view transfer and reconstruction. This integrated capability could enable valuable, reproducible stress tests.

  2. The broad comparison reveals actionable failure patterns. Table 1 covers camera-only, camera-plus-LiDAR, VLA, and world-model planners; Figures 13–15 illustrate that recognizing a hazard or predicting a plausible future does not guarantee a safe trajectory. The separation of hazard response, recovery, and rule compliance also prevents the concealment of distinct capabilities within a single aggregate score. These diagnostics are useful even if the larger causal claims remain unresolved.

Critical weaknesses
  1. Unclear added value over existing benchmarks. Tables 3–4 primarily provide binary capability checklists. The paper does not systematically identify gaps in existing scenario coverage or demonstrate that CornerBench reveals additional planner failures or diagnostic insights. Its distinctive evaluation contribution relative to Fail2Drive, WOD-E2E, and nuReasoning therefore remains insufficiently established.

  2. Insufficient validation of the claimed scoring and cost advantages. The proposed severity-aware metrics are not compared with existing metrics on the same trajectories to demonstrate meaningful additional discrimination. Likewise, the claimed economic advantage lacks comparative cost measurements accounting for generation and human filtering.

Questions for authors
  1. What specific evaluation gap does CornerBench address that existing corner-case benchmarks do not adequately cover? Please identify concrete differences in scenario coverage or evaluation capabilities beyond the binary attributes in Table 4.

  2. What comparative evidence shows that CornerBench reveals additional planner failure modes or changes conclusions about planner robustness relative to existing benchmarks? A matched comparison using representative planners would help establish its additional diagnostic value.

  3. Do the proposed severity-aware metrics provide meaningful additional discrimination when compared with existing metrics on the same trajectories? What is the construction cost per accepted case, including generation and human filtering, supporting the claimed economic advantage?

5 comments
AuthorsOpenReview ↗

Response to Reviewer jxJo (Part 1)

We sincerely appreciate your insightful questions and we believe they hold significant value for our work. We try to resolve your concerns below.

Q1: Concrete evaluation gap and relation to existing benchmarks

Thank you for raising this point. CornerBench combines controllable object insertion in recorded driving scenes with severity-aware evaluation of planner responses. This enables researchers to construct targeted hazard scenarios and distinguish differences in impact speed even when planners receive the same zero PDMS.

(1) Targeted hazard scenarios through object insertion. CornerFactory inserts uncommon obstacles, such as shipping containers, luggage trolleys, and hay bales, into recorded driving scenes. Researchers can specify the object's category, dimensions, position, and motion, allowing a desired hazard to be constructed even when it was absent from the original recording. For example, placing an obstacle on the ego path creates a direct test of braking and path adjustment; changing its position or timing allows different hazard configurations to be examined. Multi-view-consistent camera generation and LiDAR updates provide inputs for planners with different sensor interfaces. Vehicle-configuration transfer further adapts the camera observations to a target vehicle's configuration while retaining the corner-case geometry and traffic context, enabling planners from different vehicle platforms to be evaluated on equivalent cases.

(2) Graded evaluation of collision avoidance and mitigation. In challenging corner cases, collision avoidance is particularly difficult, and planners may collide despite responding differently. Under the existing PDMS metric used in NAVSIM, a zero score on the at-fault collision component makes the overall score zero, obscuring these differences. Hazard prioritizes collision avoidance while also crediting lower impact speed when a collision occurs. For example, in the luggage-trolley case, Drive-JEPA and DiffusionDrive both receive zero PDMS, but their impact speeds are 2.86 versus 8.69 m/s, yielding Hazard scores of 0.5143 versus 0.3393. Thus, the evaluation distinguishes degrees of collision mitigation that the shared zero score obscures. Q2 examines the underlying planner responses, and Q3(a) provides the same-trajectory metric comparison.

Fail2Drive provides controlled scenario construction and paired closed-loop evaluation in CARLA [R1]. WOD-E2E evaluates real-world long-tail clips using human trajectory preferences [R2], while nuReasoning provides spatial, decision, and counterfactual reasoning annotations for real-world long-tail clips [R3]. Similar hazard types can in principle be obtained through these routes. CornerBench enables researchers to construct specified hazards in recorded driving scenes and evaluate planners across vehicle configurations. Its graded scoring further distinguishes impact-speed differences that identical zero PDMS scores would obscure.

AuthorsOpenReview ↗

Response to Reviewer jxJo (Part 2)

Q2: Additional diagnostic evidence across benchmarks

Thank you for requesting concrete diagnostic evidence. Our additional analyses identify specific response weaknesses behind the scores: insufficient early slowing in hazard scenes and poor recovery from perturbed ego states. These findings help explain what a planner can and cannot handle beyond its aggregate score.

(1) Early response to hazards. On all 189 original Hazard cases, we analyze predicted trajectories before controller execution. The representative planners below span camera–LiDAR, fusion, VLA, and world-model approaches, alongside Ego-MLP. Speed reduction compares the planned speed over 0.5–1.0 s with the observed initial speed:

PlannerHazard ↑Mean speed reduction at 1 s (m/s) ↑Slowing ≥1 m/s (%) ↑
Ego-MLP0.4360.147.4
DiffusionDrive0.4320.236.9
TransFuser0.4380.2712.7
ReCogDrive-IL-2B0.4300.268.5
ReCogDrive-RL-2B0.4240.3520.1
Drive-JEPA0.4751.3342.9
SimScale0.4751.9461.9

DiffusionDrive, TransFuser, and the two ReCogDrive variants plan only 0.23–0.35 m/s of mean slowing at 1 s, with reductions of at least 1 m/s in 6.9–20.1% of cases. Drive-JEPA and SimScale plan stronger responses, at 1.33 and 1.94 m/s. These differences are present before scoring and give a behavioral explanation for the observed Hazard scores.

(2) Recovery from perturbed initial states. We also expanded the ego-state evaluation from 17 to 100 cases. The lane-keeping scores are:

PlannerEgo-state score: original 17 cases ↑Ego-state score: new 100 cases ↑
SimScale0.27410.3062
Drive-JEPA-PF0.06000.0665
Qwen-Drive RL0.11350.0622
Ego-MLP0.01410.0595
Qwen-Drive SFT0.10290.0478
LTF0.02290.0298

SimScale retains a clear lead, scoring 0.3062 on the new cohort, while the other five planners score 0.0298–0.0665. This exposes recovery from lane-straddling and off-drivable-area starts as a distinct challenge. The exported trajectories, simulated ego-vehicle states, and camera/BEV visualizations support inspection of the responses underlying these results.

Taken together, these analyses identify specific weaknesses in the evaluated planners, including limited early slowing in hazard scenes and difficulty recovering from perturbed initial states. They help explain the behaviors behind the scores and indicate which capabilities need improvement.

AuthorsOpenReview ↗

Response to Reviewer jxJo (Part 3)

Q3(a): Additional discrimination on the same trajectories

Thank you for this suggestion. We apply the existing collision criterion and Hazard to the same saved simulated trajectories, keeping the scenes, predicted trajectories, and vehicle simulation fixed. This directly tests whether the scoring rule reveals differences that a collision-only outcome obscures. Hazard reveals impact-speed differences hidden by identical collision scores on the same trajectories.

Same-trajectory comparisons. In the following three newly generated cases, Drive-JEPA and DiffusionDrive both receive zero under NAVSIM's collision criterion, resulting in zero PDMS. Their impact speeds nevertheless differ substantially. Values are shown as Drive-JEPA / DiffusionDrive:

CaseNAVSIM collision score ↑Impact speed (m/s) ↓Hazard ↑
Shipping container0 / 04.33 / 10.180.4702 / 0.2945
Luggage trolley0 / 02.86 / 8.690.5143 / 0.3393
Hay bale0 / 04.04 / 9.550.4789 / 0.3135

Appendix I, Figure 25 (p. 27) of the revised manuscript shows the predicted trajectories and simulated speed curves for these three cases.

Hazard assigns 1 when there is no at-fault collision; otherwise, H = 0.6 × (1 − clip(v / 20, 0, 1)), where v is impact ego speed in m/s. Avoiding collision therefore receives more credit than any colliding outcome, while each 1 m/s reduction in impact speed adds 0.03 within the unsaturated range. The examples show that Hazard retains meaningful differences in collision mitigation when the collision score is tied. Across the 850 cases where SimScale and Drive-JEPA both receive zero collision scores, their mean impact speeds are 6.66 and 7.35 m/s, respectively.

Overall, Hazard reveals differences in collision mitigation that are hidden by identical collision scores, while preserving the priority of collision avoidance. Evaluating the same saved trajectories shows that its added discrimination comes from recognizing differences in impact speed, enabling a more informative comparison of planner responses.

Reviewer_PTMcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

In this paper, the authors propose an automated pipeline, namely CornerFactory, for constructing corner cases based on driving logs. Specifically, by combining video-based rare-object insertion and ego-pose perturbation, CornerFactory generates multimodal driving corner cases with gaussian-based scene reconstruction, generative multi-view completion, and LiDAR ray tracing. Building upon CornerFactory, the authors further introduce a planning benchmark for autonomous driving called CornerBench, which is able to evaluate planners along various dimensions, such as robustness, understanding and reasoning. However, I still have several concerns, as detailed below.

Critical strengths
  1. The paper investigates a practical paradigm, i.e., addressing the scarcity of rare and safety-critical driving scenarios through generative scene construction.

  2. Beyond introducing CornerFactory for scenario generation, the paper further constructs CornerBench and evaluates a broad range of planners, which benefits the community.

  3. The paper is well-organized and provides sufficient visualizations, making it easy to follow.

Critical weaknesses
  1. In this paper, the authors only provide qualitative visualizations rather than any quantitative validation for CornerFactory.

    (i) The evaluation of CornerFactory is limited to qualitative comparisons in Appendix A and capability comparisons in Appendix C. Quantitative results on public evaluation data are unavailable, leaving its effectiveness and superiority over existing methods insufficiently demonstrated.

    (ii) Existing approaches either provide the metrics of image/video quality[1] or offer the metrics of downstream task performance[2] in related fields.

  2. CornerBench contains only 230 cases derived from 125 distinct source scenes, including only 17 ego-state cases and 16 red-light cases, with some scenario types represented by fewer than 10 examples. This limited scale raises concerns about whether CornerBench can reliably assess different aspects of planners and support stable rankings. Consequently, although the experiments reveal failure cases, they provide insufficient evidence to determine the contributions of different factors or identify the underlying reasons for failures.

  3. Particularly, the Hazard score is designed to distinguish failures of different severity, while 28 of the 36 models score within 0.424 to 0.444. Additionally, a large number of planners perform similarly to Ego-MLP. Even if the concentration does not imply that the metric is ineffective, it has not sufficiently demonstrated that it captures stable and meaningful differences in driving behavior.

Reference

[1] Epona: Autoregressive Diffusion World Model for Autonomous Driving. ICCV, 2025.

[2] DriveGEN: Generalized and Robust 3D Detection in Driving via Controllable Text-to-Image Diffusion Generation. CVPR, 2025.

Questions for authors

It is suggested that the authors provide further clarification and supporting evidence to address the weaknesses discussed above. In particular, the effectiveness of CornerFactory (Weakness 1) and the metric interpretations (Weaknesses 2 and 3) require further discussion.

6 comments
AuthorsOpenReview ↗

Response to Reviewer PTMc (Part 1)

We sincerely appreciate your insightful questions and we believe they hold significant value for our work. We try to resolve your concerns below.

W1: Quantitative validation of CornerFactory

Thank you for this valuable suggestion. Quantitative validation is essential because the reliability of generated scenarios directly affects the credibility of downstream planner evaluation. We therefore evaluate CornerFactory’s multi-view-consistent object insertion and free-view synthesis from perturbed ego poses. We organize the additional evaluations around these two components, assessing cross-view object consistency for the former and synthesized video quality for the latter. NeoObj improves both object-region correspondence metrics, and NeoView leads in imaging quality and subject consistency in our comparisons.

(1) Multi-view-consistent object insertion.

Evaluation protocol. Following WorldLens [R1], we assess cross-view object consistency using LoFTR [R2]. We first obtain masks of the inserted object in the front and side views using SAM 3 [R3] and visually exclude clearly non-covisible regions (to make the comparison fairer by reducing the influence of viewpoint-dependent occlusion and differences in camera overlap). We retain LoFTR matches whose endpoints lie within these object regions and report the mean match count and matching confidence across cases. The revised manuscript illustrates this protocol in Appendix G, Figure 22 (p. 24).

Comparison methods. We compare NeoObj with NVIDIA’s Cosmos-Predict1 view-extension model [R4] and NeoVerse [R5], using the same front-view reference. Cosmos-Predict1 is designed to generate mutually consistent camera views from a front-view video. NeoVerse is the foundation model for NeoObj; we fine-tune its task-specific control branch for multi-view-consistent object editing (see Appendix A.1 for training details).

Results. The target-region correspondence results are:

Target-region measurementCosmos-Predict1NeoVerseNeoObj
Mean LoFTR matches ↑141.2146.9164.8
Mean matching confidence ↑0.5510.6100.653

NeoObj achieves higher match counts and matching confidence than both baselines within the evaluated object regions. This ordering is consistent with the reconstruction-guided generation paradigm of NeoVerse and NeoObj: explicit 3D reconstruction and cross-view rendering provide geometric and appearance cues for preserving the inserted object across cameras. These cues offer a plausible explanation for their higher correspondence scores than Cosmos-Predict1, which performs front-video-conditioned multi-view generation. NeoObj further improves over NeoVerse by 17.9 matches and 0.043 confidence, consistent with the benefit of task-specific fine-tuning to complete transferred objects and correct localized cross-view artifacts (Appendix A.1). Qualitative comparisons of the full camera views and enlarged object regions are provided in Appendix G, Figure 23 (p. 25) of the revised manuscript.

(2) Free-view synthesis from perturbed ego poses.

Evaluation protocol. We evaluate 50 distinct source scenes from 27 driving logs, with three cameras per scene (150 videos per method), using three VBench [R6] dimensions: Imaging Quality (MUSIQ), Subject Consistency, and Motion Smoothness. We score the same first 16 frames at 10 Hz on a shared 560×224 image region, matching FreeVS's native crop. Scores average cameras within each scene and then scenes; confidence intervals in the accompanying evidence are clustered by driving log.

Comparison methods. We compare NeoView with TrajectoryCrafter [R7], NeoVerse [R5], and the driving-specific method FreeVS [R8], using the same scenes and target camera poses.

Results. The video-quality results are:

MethodImaging Quality (MUSIQ)Subject ConsistencyMotion Smoothness
TrajectoryCrafter0.48980.90380.9785
NeoVerse0.46900.91320.9765
FreeVS0.51610.91460.9742
NeoView0.55010.93400.9778

NeoView achieves the highest Imaging Quality and Subject Consistency. Its Motion Smoothness is higher than NeoVerse and FreeVS and close to TrajectoryCrafter.

These gains are consistent with NeoView's task-specific training on multi-camera driving videos (Appendix A.2). The model is trained to complete missing regions and correct appearance discontinuities and fragmented structures in reconstructed views. This adaptation helps explain the improved visual quality and subject consistency when synthesizing observations from perturbed ego poses. Visual comparisons at matched target cameras and times are provided in Appendix H, Figure 24 (p. 26) of the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer PTMc (Part 2)

W2: Dataset scale and ranking stability

Thank you for highlighting this important issue. Dataset size is crucial because sufficient samples are needed to assess planner robustness reliably across diverse corner cases. We have therefore increased the number of cases and examined ranking stability on the enlarged dataset. On the enlarged evaluation set, the same four planners lead on Hazard, and SimScale remains first on ego-state recovery.

(1) Increased dataset size and diversity.

Case categoryOriginal setExpanded experimental setScale increase
Total2301,3005.7×
Object / scene editing2131,2005.6×
Ego-state perturbation171005.9×

Beyond increasing the number of cases, we also broaden object diversity: the original uncommon-object subset covers 7 object types, while the expanded insertion set covers 100 object categories, spanning animals, vehicles, furniture, household appliances, construction equipment, traffic facilities, and road obstacles.

Public access to the expanded object set. We have anonymously released all 1,200 object / scene editing cases in CornerBench-1200, together with an interactive three-camera gallery. Each case includes front-left, front, and front-right videos at 1280×720; the gallery supports synchronized playback and search across all released cases.

(2) Hazard ranking stability. We compare the 34 planners evaluated on both datasets, showing their original Table 1 order and expanded-set ranks within this common group.

PlannerRank ↓ (old → new)Original Hazard ↑Expanded Hazard ↑95% interval
SimScale1 → 1 (=)0.4750.5117[0.4881, 0.5355]
GTRS-Aug3 → 2 (↑1)0.4680.5095[0.4814, 0.5386]
Drive-JEPA2 → 3 (↓1)0.4750.5083[0.4793, 0.5360]
GTRS-Dense4 → 4 (=)0.4610.4983[0.4734, 0.5222]
SparseDriveV226 → 5 (↑21)0.4340.4975[0.4752, 0.5202]
DrivoR8 → 6 (↑2)0.4430.4918[0.4678, 0.5153]
DriveWorld-VLA27 → 7 (↑20)0.4330.4903[0.4672, 0.5146]
RAP17 → 8 (↑9)0.4380.4890[0.4690, 0.5093]
VeteranAD6 → 9 (↓3)0.4520.4838[0.4589, 0.5088]
WoTE13 → 10 (↑3)0.4400.4800[0.4543, 0.5056]
Hydra-MDP5 → 11 (↓6)0.4540.4800[0.4544, 0.5057]
AutoVLA20 → 12 (↑8)0.4360.4766[0.4547, 0.4993]
SeerDrive9 → 13 (↓4)0.4410.4718[0.4449, 0.4982]
Ego-MLP21 → 14 (↑7)0.4360.4715[0.4485, 0.4947]
DiffusionDriveV223 → 15 (↑8)0.4340.4683[0.4434, 0.4925]
DriveLaW7 → 16 (↓9)0.4440.4644[0.4453, 0.4837]
ReCogDrive-RL-2B34 → 17 (↑17)0.4240.4632[0.4414, 0.4848]
ReCogDrive-RL-8B33 → 18 (↑15)0.4240.4607[0.4407, 0.4805]
WAM-Flow15 → 19 (↓4)0.4390.4605[0.4384, 0.4822]
PRIX32 → 20 (↑12)0.4290.4603[0.4391, 0.4815]
DynVLA12 → 21 (↓9)0.4410.4603[0.4420, 0.4781]
LEAD25 → 22 (↑3)0.4340.4565[0.4340, 0.4791]
Drive-JEPA-PF18 → 23 (↓5)0.4370.4559[0.4332, 0.4780]
GoalFlow22 → 24 (↓2)0.4350.4556[0.4322, 0.4789]
TransFuser16 → 25 (↓9)0.4380.4555[0.4337, 0.4766]
CuriousVLA14 → 26 (↓12)0.4400.4537[0.4308, 0.4764]
MeanFuser19 → 27 (↓8)0.4360.4530[0.4308, 0.4748]
Mimir24 → 28 (↓4)0.4340.4520[0.4284, 0.4753]
OneVL11 → 29 (↓18)0.4410.4508[0.4310, 0.4700]
LTF29 → 30 (↓1)0.4320.4482[0.4287, 0.4675]
DiffusionDrive28 → 31 (↓3)0.4320.4471[0.4247, 0.4696]
DriveVLA-W010 → 32 (↓22)0.4410.4468[0.4266, 0.4665]
ReCogDrive-IL-8B30 → 33 (↓3)0.4300.4414[0.4215, 0.4606]
ReCogDrive-IL-2B31 → 34 (↓3)0.4300.4335[0.4108, 0.4556]

SimScale, Drive-JEPA, GTRS-Aug, and GTRS-Dense remain the top four, supporting the stability of the leading group as dataset size and object diversity increase. In particular, the continued strong performance of Drive-JEPA and SimScale is consistent with our training-data analysis in Section 5.1: Drive-JEPA learns from driving videos covering diverse long-tail events, while SimScale trains on synthetic off-nominal driving situations, highlighting the importance of diverse corner-case training data for planner robustness.

AuthorsOpenReview ↗

Response to Reviewer PTMc (Part 3)

(3) Ego-state ranking stability. We compare all six evaluated planners on the original 17 cases and the new 100 cases using the Ego-state (lane-keeping) score:

PlannerRank ↓ (old → new)Original score (17 cases) ↑New score (100 cases) ↑
SimScale1 → 1 (=)0.27410.3062
Drive-JEPA-PF4 → 2 (↑2)0.06000.0665
Qwen-Drive RL2 → 3 (↓1)0.11350.0622
Ego-MLP6 → 4 (↑2)0.01410.0595
Qwen-Drive SFT3 → 5 (↓2)0.10290.0478
LTF5 → 6 (↓1)0.02290.0298

SimScale remains first by a clear margin, scoring 0.2741 on the original 17 cases and 0.3062 on the new 100 cases. This is consistent with our analysis in Section 5.1 and Figure 10: SimScale’s training data combines perturbed ego states with trajectories that return toward the lane center, providing explicit recovery supervision. Its continued lead on the larger cohort strengthens the evidence for its advantage in recovering from lane-straddling and off-drivable-area starts.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Paper revision uploaded and looking forward to further discussion

Dear Area Chair and Reviewers,

We sincerely appreciate the time and effort you have devoted to reviewing our work. Your constructive suggestions have helped us improve the manuscript and strengthen our evaluation. We are also grateful for your recognition of the following strengths of our work.

  • Constructing rare hazards in real driving scenes. Reviewer PTMc highlighted the practical value of addressing the scarcity of safety-critical events in driving logs. CornerFactory lets researchers specify an obstacle’s type, position, and motion, then introduce it into a recorded scene to test a planner’s response to a particular hazard.

  • Comparing planners on shared cases. Reviewers jxJo and 8mcc highlighted the value of consistent camera views and aligned LiDAR. These capabilities allow planners with different sensor configurations to encounter the same hazard, supporting the reproducible stress tests noted by Reviewer jxJo.

  • Revealing differences in driving behavior. Reviewer jxJo highlighted the actionable failure patterns revealed by our planner comparison, including cases where a planner recognizes a hazard yet still plans an unsafe trajectory. Reviewer 8mcc also recognized the value of severity-aware evaluation, which distinguishes substantial braking before impact from little or no braking, even when both planners collide.

In response to your feedback, we made the following updates.

  • Expanded evaluation and anonymous public release. We expanded the evaluation from 230 to 1,300 cases, comprising 1,200 object and scene editing cases and 100 ego-state perturbation cases. We have anonymously released all 1,200 object and scene editing cases for public access in the CornerBench-1200 dataset.

  • Ranking stability on the expanded evaluation. To address Reviewer PTMc’s concerns about the limited evaluation scale and whether it supports reliable planner rankings, we compared the 34 planners evaluated on both sets. The same four planners ranked highest on hazard response in both evaluations. SimScale also retained its lead in ego-state recovery as that evaluation grew from 17 to 100 cases.

  • Quantitative evaluation of generation quality. Appendices G and H add quantitative comparisons of cross-camera object consistency and view synthesis from perturbed ego poses, providing additional evidence for the quality of the generated observations.

  • Planner behavior and metric effectiveness. Appendix I adds trajectory and speed analyses alongside metric comparisons on the same predicted trajectories, providing evidence that our severity-aware scores reflect differences in planner behavior. In one luggage-trolley scenario, two planners both received zero PDMS, while their impact speeds were 2.86 m/s and 8.69 m/s, corresponding to Hazard scores of 0.5143 and 0.3393. The speed profiles show how much each planner slowed before impact, supporting the effectiveness of Hazard in distinguishing meaningful differences in collision mitigation.

According to the official reviewing schedule, the author–reviewer discussion period concludes on November 18, followed by private discussion among reviewers and area chairs. We would greatly appreciate any further feedback on our responses and additional results during the remaining discussion period. We look forward to continuing the discussion with the reviewers, addressing any outstanding concerns, and further improving the quality of our work.

Best regards,
The Authors

Dear reviewers,

As the rebuttal period is coming to an end, I kindly remind you to check the authors' responses. If you have not yet responded, please indicate whether your concerns have been adequately addressed. Many thanks!

Best, AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.