Benchmarking end-to-end planners in corner cases
Is your driving model safe enough?
Build the corner cases that real-world logs rarely capture. Evaluate driving models with continuous, severity-aware metrics that capture graded safety differences.
1 Institute of Automation, Chinese Academy of Sciences2 Li Auto Inc.* Corresponding authors † Project lead
Create targeted object-centered corner cases by inserting a specified rare object or abnormal road user along an intended 3D trajectory. These cases test whether multi-view planners can recognize the hazard and slow down, stop, or steer away safely.
Original / EditedCreate ego-state corner cases from off-nominal poses, such as lane-straddling or non-drivable-area starts, while preserving the original map, route, and surrounding traffic. These cases test whether a planner actively recovers toward the lane center or another safer nominal state.
Original / Perturbed poseTransfer the same generated corner case across vehicle platforms and camera configurations while preserving its geometry and traffic context. This enables planners built for different visual input formats to be compared on equivalent scenarios and tests whether they respond safely to the same hazard.
Source / Target rigCamera-based editing alone is insufficient for multimodal planners because an inserted object would appear in the RGB observations but remain absent from the original LiDAR point cloud. Aligning the LiDAR returns with the edited camera views enables the generated corner cases to evaluate planners that require both camera and LiDAR inputs.
Drag to rotate · scroll to zoom
CornerBench organizes generated corner cases into a structured taxonomy and evaluates planner responses with continuous, severity-aware metrics. The benchmark contains 230 cases across 19 scenario types, constructed from 125 nuPlan source scenes.
Using the pipeline described above, we construct a suite of meaningful corner cases that are challenging for driving and informative for evaluating current planning models. CornerBench organizes generated cases into three major groups: Ego-state cases, such as a lane-straddling start; Environmental cases, such as smoke; and Object-centric cases. Object-centric cases are further divided according to whether the challenge comes from an uncommon object, such as a fallen bicycle; uncommon behavior of a common object, such as a red-light runner; or uncommon texture of a common object, such as a bus with a stop-sign marking.
Binary pass/fail criteria collapse different failures into the same zero. CornerBench retains the original success condition while assigning graded scores to failures according to collision speed, lane-deviation duration, or red-light overlap.
Distinguishes low-speed impact from driving through a hazard at high speed.
Grades how long the trajectory remains outside the lane-center tolerance.
Separates minor boundary intrusion from entering deeply into a forbidden region.
clip(x) = min(max(x, 0), 1)
Select any endpoint on the road. The start and 4-second planning horizon remain fixed; this demo uses vmax = 15 m/s.
Sorted by Hazard-response score. Y marks a used capability or additional data source. VLA: vision-language-action; WM: world model; Extra: data beyond NAVSIM navtrain; Sens.: C, camera; C+L, camera and LiDAR; N, ego status only.
| Planner | VLA | WM | Extra | Sens. | Hazard response | Lane keeping | Traffic-rule compliance |
|---|---|---|---|---|---|---|---|
| Alpamayo R1 | Y | - | Y | C | 0.4761 | 0.081 | 0.7403 |
| SimScale | - | - | Y | C | 0.4752 | 0.2741 | 0.312 |
| Drive-JEPA | - | Y | Y | C | 0.4753 | 0.1483 | 0.084 |
| GTRS-Aug | - | - | - | C | 0.468 | 0.122 | 0.133 |
| GTRS-Dense | - | - | - | C | 0.461 | 0.146 | 0.395 |
| Alpamayo 1.5 | Y | - | Y | C | 0.457 | 0.143 | 0.8902 |
| Hydra-MDP | - | - | - | C | 0.454 | 0.1772 | 0.304 |
| VeteranAD | - | - | - | C | 0.452 | 0.051 | 0.128 |
| DriveLaW | - | Y | Y | C | 0.444 | 0.059 | 0.292 |
| DrivoR | - | - | - | C | 0.443 | 0.139 | 0.181 |
| SeerDrive | - | Y | - | C+L | 0.441 | 0.138 | 0.277 |
| DriveVLA-W0 | Y | Y | - | C | 0.441 | 0.059 | 0.166 |
| OneVL | Y | Y | - | C | 0.441 | 0.085 | 0.506 |
| DynVLA | Y | Y | - | C | 0.441 | 0.080 | 0.246 |
| WoTE | - | Y | - | C+L | 0.440 | 0.088 | 0.228 |
| CuriousVLA | Y | - | - | C | 0.440 | 0.101 | 0.210 |
| WAM-Flow | Y | - | Y | C | 0.439 | 0.094 | 0.9231 |
| TransFuser | - | - | - | C+L | 0.438 | 0.046 | 0.240 |
| RAP | - | - | Y | C | 0.438 | 0.109 | 0.016 |
| Drive-JEPA-PF | - | Y | Y | C | 0.437 | 0.060 | 0.253 |
| MeanFuser | - | - | - | C | 0.436 | 0.099 | 0.283 |
| AutoVLA | Y | - | - | C | 0.436 | 0.000 | 0.442 |
| Ego-MLP | - | - | - | N | 0.436 | 0.014 | 0.217 |
| GoalFlow | - | - | - | C+L | 0.435 | 0.062 | 0.182 |
| DiffusionDriveV2 | - | - | - | C+L | 0.434 | 0.120 | 0.062 |
| Mimir | - | - | - | C+L | 0.434 | 0.129 | 0.198 |
| LEAD | - | - | - | C | 0.434 | 0.071 | 0.312 |
| SparseDriveV2 | - | - | - | C | 0.434 | 0.104 | 0.040 |
| DriveWorld-VLA | Y | Y | - | C | 0.433 | 0.129 | 0.284 |
| DiffusionDrive | - | - | - | C+L | 0.432 | 0.030 | 0.214 |
| LTF | - | - | - | C | 0.432 | 0.023 | 0.337 |
| ReCogDrive-IL-8B | Y | - | - | C | 0.430 | 0.064 | 0.361 |
| ReCogDrive-IL-2B | Y | - | - | C | 0.430 | 0.098 | 0.427 |
| PRIX | - | - | - | C | 0.429 | 0.051 | 0.279 |
| ReCogDrive-RL-8B | Y | - | - | C | 0.424 | 0.099 | 0.038 |
| ReCogDrive-RL-2B | Y | - | - | C | 0.424 | 0.128 | 0.166 |
Training-data scale. Alpamayo R1 and Alpamayo 1.5 perform strongly with large-scale external driving data and without using the NAVSIM navtrain set, showing strong cross-domain generalization. This suggests that increasing training-data volume can bring clear gains to planning models. However, data volume alone does not explain the quantitative results.
Drive-JEPA, SimScale, DriveLaW, and WAM-Flow use additional data at comparable scales. Nevertheless, Drive-JEPA and SimScale achieve Hazard-response scores of 0.475 and 0.475, respectively, compared with 0.444 for DriveLaW and 0.439 for WAM-Flow. A closer inspection reveals a key difference in data composition. DriveLaW and WAM-Flow rely primarily on standard driving data from nuPlan and nuScenes. By contrast, Drive-JEPA uses driving videos covering diverse long-tail events, while SimScale perturbs ego trajectories in ordinary driving scenes to synthesize diverse off-nominal situations.
Drive-JEPA pretrains its V-JEPA backbone on 208 hours of driving video assembled from CoVLA, DrivingDojo, and other datasets before training the downstream NAVSIM planner. This collection exposes the visual encoder to a broader range of object interactions than NAVSIM-style planning data alone. The examples below include overhanging branches, animals near the ego lane, and nighttime roadside fire. This diversity is reflected in cases where Drive-JEPA produces a more conservative response to the corner-case hazard than other planners.
SimScale augments its training data with 384K synthetic off-nominal scenes. It perturbs ego trajectories in ordinary driving scenes and uses rule-based strategies or existing planners to generate trajectories that return the vehicle toward normal lane-centered states, providing explicit supervision for ego-state recovery. This targeted augmentation is reflected in the highest Lane-keeping score of 0.274 and in the cases below, where SimScale steers back toward the lane center while most planners continue along the perturbed initial pose.
For Drive-JEPA, driving-video pretraining on long-tail events is associated with stronger responses to unusual objects. For SimScale, training with synthetic off-nominal states and recovery trajectories is associated with clearer lane-centering behavior. Together, these cases suggest that covering rare objects and off-nominal ego states is more important than merely increasing the volume of routine driving data.
Model size. The model-size comparison shows no clear positive relationship between reported model size and CornerBench Hazard-response score. Within the same ReCogDrive family, scaling the backbone from 2B to 8B leaves the score unchanged in both settings, at 0.430 for IL and 0.424 for RL, while models with smaller active planners such as Drive-JEPA and SimScale remain among the strongest methods. Simply increasing active model size is therefore not sufficient; robustness also depends on data diversity.
Answer Training-data diversity is more important for corner-case robustness than simply increasing data volume or model size.
Many evaluated planners use LiDAR together with camera inputs to obtain more reliable geometric perception of object distances and spatial structure. However, the quantitative ranking shows that several top-performing planners do not use a LiDAR sensor.
The qualitative results below show three CornerBench test scenes. DiffusionDrive receives a 2D LiDAR BEV input, whereas Ego-Status MLP uses no LiDAR and does not even use camera input; nevertheless, the trajectories from both planners fail to avoid the frontal obstacle. This indicates that adding a LiDAR sensor does not by itself ensure stronger corner-case response.
Answer Adding a LiDAR sensor does not by itself ensure stronger corner-case response.
VLA planners leverage the strong scene-understanding capability of VLMs to incorporate visual observations and language-level representations into action planning. In the examples below, ReCogDrive's VLM output identifies the inserted hazard or red traffic light and proposes a reasonable high-level response, but the final trajectory remains inconsistent with that response.
In the horse-crossing case, for example, the VLM output first recognizes the horse crossing ahead and then recommends braking to a complete stop, but the planner's trajectory still collides with the horse. This shows that VLM scene understanding is not reliably converted into correct trajectory-planning output.
Answer VLMs can show strong scene understanding, but VLA planners do not always convert that understanding into safe trajectory plans.
CoT is widely regarded as a promising way to enhance the reasoning and planning capabilities of VLA planners in complex driving situations. We test four CoT-enabled VLA planners on the same corner-case scenes.
OneVL states that there are no relevant objects, CuriousVLA marks the critical object category as absent, and AutoVLA treats the scene as a straightforward driving case. Alpamayo 1.5 instead recognizes the corner-case hazard and suggests a safer response. The first three planners are trained primarily on regular driving scenes, whereas Alpamayo 1.5 uses 80,000 hours of diverse driving data. This comparison suggests that CoT alone does not ensure generalization to corner cases; its effectiveness also depends strongly on training-data diversity.
Answer CoT planners trained only on regular driving scenes may fail to generalize to corner cases.
World models are widely expected to improve driving planning by predicting how a scene may evolve and using the predicted future to inform planning. To examine whether this benefit extends to corner-case scenarios, we inspect the future images decoded by OneVL together with its corresponding planned trajectories on CornerBench.
The decoded futures appear visually plausible, but neither the decoded futures nor the corresponding trajectories show the correct response to a frontal obstacle or a Lane-Straddling Start case. One likely reason is that OneVL is trained mainly on regular driving scenes, so its decoded futures do not generalize reliably to these rare corner-case situations.
Answer A world model does not by itself ensure a correct corner-case response: although the decoded future images may appear visually plausible, neither the predicted scene evolution nor the corresponding planned trajectory necessarily reflects the correct response.
04 / Citation
Paper identifiers will be updated with the public release.
@misc{shi2026cornerbench,
title = {CornerBench: Is Your Driving Model Safe Enough?},
author = {Shi, Ziqi and Yang, Yuxue and Li, Yanjia and
Shao, Maoyuan and Wang, Yida and Zhang, Xueyang and
Zhan, Kun and Fan, Lue and Zhang, Zhaoxiang},
year = {2026}
}