Literature
Display
← All papers

Challenger: Affordable Adversarial Driving Video Generation for Safety Testing

Zhiyuan Xu · Bohan Li · Huan-ang Gao · Mingju Gao · Xin Jin · Hang Zhao · Shuo Feng · Ya-Qin Zhang · Hao Zhao

· IASEAI 2026 ↗

Robotics9 min text · CondensedOriginal paper ↗
Contents

Overview

Challenger generates adversarial traffic trajectories and renders them into six-camera driving videos for testing vision-based planners.

Refer to caption
Figure 1: Photorealistic adversarial driving videos generated by Challenger. Each scenario includes an adversarial vehicle—highlighted with a white 3D bounding box in camera views and depicted as a red rectangle in the bird’s-eye view (BEV) map—that is intentionally designed to challenge the ego vehicle through aggressive or unexpected maneuvers. Challenger autonomously produces these videos using a diffusion-based trajectory generator, physics-aware planning, and multiview neural rendering, with an affordable computation budget. Readers are suggested to zoom in on BEV maps and camera views for detailed inspection. Videos are available at our project page

At a glance

Challenger combines a trajectory diffusion model, kinematic simulation, a geometric challenge score, and multiview video generation. It searches over traffic motion before paying the cost of rendering, then tests four pretrained autonomous-driving planners on the resulting Adv-nuSc dataset. Predicted collisions rise substantially, and failure-selected scenarios often affect multiple architectures. The evidence concerns open-loop three-second plans; it does not measure closed-loop driving or physical crash rates.

1 Introduction

Naturalistic driving datasets contain relatively few deliberately difficult traffic interactions. Adversarial trajectory generators and photorealistic video generators address different parts of this testing problem. Challenger joins them: it changes a background vehicle’s motion within a recorded scene, rejects physically unsuitable or already-colliding trajectories, and synthesizes camera observations for the remaining interaction. Its contribution is an automated sensor-level stress-test generator, not a deployed driving policy or a certified safety assessment.

2 Related Work

Prior work provides adversarial traffic generation, controllable driving imagery, and end-to-end planning architectures; Challenger integrates these capabilities in an offline testing pipeline.

2.1 Driving Trajectory for Planning

Naturalistic adversarial testing, rare-event simulation, diffusion planning, and language-conditioned traffic generation improve coverage of unusual behaviors. Challenger draws its resample-and-denoise search from Diffusion-ES [14], adding kinematic feasibility checks and a renderer-compatible geometric score. Unlike optimization against every rendered candidate, this search can proceed without querying a victim driving model.

2.2 Driving Video Generation

Driving video generators condition on scene structure to maintain geometry, motion, and camera consistency. Challenger uses MagicDriveDiT [24] as its renderer, while supplying traffic motions selected to challenge the ego vehicle. The rendering model is an existing component; the proposed contribution is the adversarial scenario-generation process around it.

2.3 End-to-End Autonomous Driving

End-to-end driving systems combine perception and planning, but their learned predictions may fail under unusual interactions. The paper evaluates official pretrained UniAD, VAD, SparseDrive, and DiffusionDrive models. Diverse architectures provide a useful transfer test, although sharing failures on this generated distribution does not establish a common internal cause.

3 Method

The pipeline samples trajectories, simulates their execution, scores geometric plausibility and proximity, repeatedly refines promising candidates, and renders only selected scenes.

Refer to caption
Figure 2: Challenger Overview. Challenger first ingests 3D bounding boxes and BEV road maps from a real-world dataset (e.g., nuScenes) to initialize a driving scene. It then randomly selects a background vehicle to act as the adversarial agent, while leaving all other participants unchanged. At fixed keyframes, Challenger plans the adversarial vehicle’s trajectory and executes it continuously between keyframes. At each planning keyframe, a multi-round refinement process is used to explore the trajectory space efficiently and generate an adversarial one. In the first round, a batch of candidate trajectories τ∈RB×T×2\tau\in R^{B\times T\times 2} is sampled from a diffusion-based generator. To ensure physical feasibility, a physics-aware planning simulator with an LQR controller and a kinematic model is used to simulate the motion of the adversarial vehicle. The simulated trajectories are then scored, and the top-performing trajectories are resampled (with replacement), perturbed by noise, denoised via the diffusion model, and fed back into the planning simulator for subsequent rounds. This iterative process continues for a fixed number of rounds in order to progressively narrow down the trajectory space in an efficient manner, after which the best simulated trajectory is selected as the final adversarial driving plan for that keyframe. The generated trajectory modifies the original driving scene, and this process is repeated across all keyframes. Finally, a multiview neural renderer synthesizes photorealistic video outputs of the final adversarial scenario.

3.1 Diffusion Trajectory Generator

An unconditional diffusion model is trained on nuPlan trajectories [66], represented by fixed-length sequences of two-dimensional waypoints in local coordinates. It can draw new motion plans from noise or perturb an existing plan by partial noising and denoising. The learned prior supplies motion diversity; because it is unconditional, environmental compatibility must be established downstream.

3.2 Physics-aware Planning Simulator

A linear-quadratic regulator tracks each candidate using steering and acceleration commands, while a kinematic bicycle model produces the resulting vehicle motion and bounding boxes. This checks a simplified execution model rather than treating sampled waypoints as automatically feasible. It does not model realistic crash dynamics, which is one reason the generator rejects collisions.

3.3 Trajectory Scorer

The score penalizes leaving drivable space and contact with other vehicles, including the recorded ego vehicle, while favoring close approaches. It operates on simulated trajectories and bounding boxes, avoiding expensive rendering and planner evaluation for every candidate. Proximity is a surrogate for difficulty, not a direct guarantee that any given driving model will fail.

3.4 Multi-round Trajectory Refinement

Each round simulates and scores a batch, resamples promising candidates with replacement, then partially noises and denoises them to explore nearby motions. The trained generator stays fixed during this search. The final scene is retained only if the adversarial vehicle avoids collisions and the original ego trajectory remains collision-free.

This establishes feasibility of the recorded reference path under the generator’s checks, not recoverability by every learned planner from synthetic observations. The paper’s main prose and algorithm select the highest score, whereas the appendix defines lower scores as better and samples with a negative-score softmax. The original algorithm is retained, including its use of the pre-resampling trajectory symbol in the noising line; these notation inconsistencies should be resolved before implementation.

3.5 Multiview Neural Renderer

MagicDriveDiT receives the modified three-dimensional boxes and bird’s-eye-view map and generates synchronized observations for six cameras. Rendering occurs after trajectory search and filtering. Re-rendering unchanged validation scenes provides a useful control for image-generation shift, though the adversarial motions can introduce a different shift from unchanged scenes.

4 Experiments

The experiments compare open-loop planning collisions, transfer across architectures, component ablations, visual quality, and generation cost.

4.1 Experiment Setup

The base nuScenes validation split contains 150 twenty-second scenes and 6,019 annotated samples. Its unchanged scenes are re-rendered as nuScenes-val-R. Challenger produces 156 adversarial scenes and 6,115 samples for Adv-nuSc. Four official pretrained vision-based models predict three-second ego trajectories from multiview images; the test evaluates these predictions against scene geometry rather than rolling the policies through an interactive closed loop.

4.2 Results on the Adv-nuSc Dataset

At three seconds, collisions rise from 0.61% to 6.96% for UniAD, 0.42% to 9.08% for VAD, 0.238% to 2.430% for SparseDrive, and 0.169% to 3.646% for DiffusionDrive. The corresponding unchanged re-rendered rates are 0.64%, 0.59%, 0.370%, and 0.403%. Thus rendering alone alters performance, but the adversarial scenarios cause much larger absolute increases.

The table’s average columns average one-, two-, and three-second rates; they are different from the three-second endpoint metric. Its annotated multiplicative increases do not uniformly match simple ratios of the displayed averages, so the direct rates are the clearest quantitative evidence. No closed-loop collision probability is established.

Dataset AD Model 1s 2s 3s avg AD Model 1s 2s 3s avg
nuScenes-val UniAD [8] 0.10% 0.15% 0.61% 0.29% SparseDrive [10] 0.020% 0.063% 0.238% 0.107%
nuScenes-val-R 0.07% 0.24% 0.64% 0.32% 0.010% 0.090% 0.370% 0.156%
Adv-nuSc 0.80% 4.10% 6.96% 3.95% 12.6×↑12.6\times\uparrow 0.029% 0.618% 2.430% 1.026% 8.6×↑8.6\times\uparrow
nuScenes-val VAD [9] 0.11% 0.24% 0.42% 0.26% DiffusionDrive [11] 0.059% 0.068% 0.169% 0.099%
nuScenes-val-R 0.19% 0.32% 0.59% 0.37% 0.030% 0.055% 0.403% 0.163%
Adv-nuSc 4.46% 7.59% 9.08% 7.05% 26.1×↑26.1\times\uparrow 0.068% 1.299% 3.646% 1.671% 15.9×↑15.9\times\uparrow
Table 1: Performance of End-to-End Autonomous Driving Systems Across Different Datasets. We report the collision rate over 3 seconds and the average collision rate for each dataset. Tested E2E AD models perform much worse on adversarial driving scenarios generated by Challenger.

4.3 Qualitative Examples of Adversarial Driving Scenarios

Two UniAD examples show an overtake from the right misread as a right turn, and a left cut-in misread as continued straight motion. The generated traffic therefore exposes mismatches between predicted and actual agent trajectories. Training-data bias is the authors’ proposed explanation, not a directly measured cause in these examples.

Refer to caption
(a) Surpassing from the right.
Refer to caption
(b) Cut-in from the left.
Figure 3: Failure cases of an E2E AD model in adversarial scenarios generated by Challenger.

4.4 Attack Transferability

For each surrogate model, retain scenes in which it predicts a collision in at least one sample, then measure how often another model also fails in at least one sample. The plotted off-diagonal scene-level rates range from 44.7% to 85.1%. These conditional transfer rates differ from the per-sample collision rates in the main table and do not describe random unfiltered scenes.

Settings Collision Rate
Challenger 6.17%
   - MTR 3.49%
   - TS 0.13%
    - MTR - TS 0.80%
Table 2: Ablation study. We show the effectiveness of the multi-round trajectory refinement process (MTR) and trajectory scoring process (TS).
Dataset SC IQ
nuScenes-val 0.8349 0.3801
nuScenes-val-R 0.8173 0.3471
Adv-nuSc 0.8093 0.3469
Figure 4: Quantitative evaluation of video quality. We report the subject consistency (SC) and imaging quality (IQ) of the datasets.
Original figure
Figure 5: Attack Transferability. This figure illustrates the failure rates of victim models when evaluated on adversarial driving scenarios filtered based on the performance of a surrogate model. A failure is recorded when the victim model collides with another object in at least one sample of a given scenario. The failure rate is computed as the number of such scenarios divided by the total number of filtered scenarios.

4.5 Ablation Study.

On twenty selected validation scenes, full Challenger produces a 6.17% failure rate for UniAD, versus 3.49% without multiround refinement, 0.13% without trajectory scoring, and 0.80% without either component. Scoring is essential in this comparison, while refinement further raises difficulty. The non-monotonic no-scoring variants and limited scene subset caution against interpreting these results as a universal independent contribution of each module.

4.6 Video Quality Evaluation

Vehicle-focused subject consistency is 0.8349 on original imagery, 0.8173 on unchanged re-rendering, and 0.8093 on Adv-nuSc. Imaging quality is 0.3801, 0.3471, and 0.3469 respectively. These diagnostics reveal a rendering gap despite convincing examples. They assess generated appearance and temporal coherence, not exact physical fidelity or a calibrated probability of real-world failure.

Refer to caption
Figure 6: Representative video frames from datasets. The re-rendered and adversarial datasets maintain photorealistic quality comparable to the original nuScenes dataset.
Refer to caption
Figure 7: Qualitative example of MTR and TS.

5 Conclusion

Searching traffic motion before rendering makes sensor-level adversarial testing practical and reveals substantial vulnerabilities in the evaluated open-loop planners. The results support using generated scenarios as additional stress tests, while leaving closed-loop behavior and transfer to real traffic unmeasured.

6 Limitations

The renderer can introduce artifacts and distribution shift. Evaluation is limited to open-loop, vision-based end-to-end systems; closed-loop interaction and additional modalities such as lidar remain future work. The simplified vehicle simulation, geometric difficulty proxy, and collision-free reference-path check also limit the safety conclusions that can be drawn.

Appendix

The appendix specifies scoring, refinement, hyperparameters, runtime, video-quality measurement, dataset construction, multiple adversarial vehicles, and further failure examples.

Appendix A Implementation Details

Algorithm 1 gives the full scene-generation loop. The appendix supplies exact scoring and sampling formulas and fixed parameter values; its score convention must be reconciled with the algorithm’s maximization notation.

Original Algorithm 1, all twenty-three lines and loop indentation
Algorithm 1 Adversarial Scene Generation.
The algorithm takes 3D bounding boxes and a BEV map from a real-world scene as input, and makes adversarial driving plans for an adversarial vehicle periodically. On each planning key frame, a multi-round trajectory refinement process is performed, and the final trajectory is used to update the 3D bounding boxes. The algorithm then renders the scene using the updated bounding boxes and BEV map.

A.1 Trajectory Scoring

Lower scores are preferred. Drivable-space penalties use the center and four corners of each vehicle box, penalizing center violations more than corner violations. Collision checks combine distance screening with box intersections. Time weights emphasize early road or collision violations, while the proximity term gives more weight to later close approaches.

The final objective weights road violations, collisions, and normalized ego distance by 5, 10, and 1, with a small random initialization. The displayed normalization divides by the within-batch distance-score range; behavior when that range is zero is not specified. These choices discourage pre-existing crashes, which neither the kinematic simulator nor the renderer is designed to represent.

Equations
DACScore\displaystyle\mathrm{DACScore} =CenterViolation⊙DACStepWeight\displaystyle=\mathrm{CenterViolation}\odot\mathrm{DACStepWeight}
+CornerWeight×CornersViolation⊙DACStepWeight\displaystyle\quad+\mathrm{CornerWeight}\times\mathrm{CornersViolation}\odot\mathrm{DACStepWeight}
CRScore=AnyCollision⊙ColStepWeight\displaystyle\mathrm{CRScore}=\mathrm{AnyCollision}\odot\mathrm{ColStepWeight}
ACScore′=\displaystyle\mathrm{ACScore^{\prime}}= Dist⊙ACStepWeight\displaystyle\mathrm{Dist}\odot\mathrm{ACStepWeight}
ACScore=\displaystyle\mathrm{ACScore}= A​C​S​c​o​r​e′−min⁡(A​C​S​c​o​r​e′)max⁡(A​C​S​c​o​r​e′)−min⁡(A​C​S​c​o​r​e′)\displaystyle\frac{ACScore^{\prime}-\min(ACScore^{\prime})}{\max(ACScore^{\prime})-\min(ACScore^{\prime})}
FinalScore=\displaystyle\mathrm{FinalScore}= DACWeight⋅DACScore+CRWeight⋅CRScore+ACWeight⋅ACScore\displaystyle\mathrm{DACWeight}\cdot\mathrm{DACScore}+\mathrm{CRWeight}\cdot\mathrm{CRScore}+\mathrm{ACWeight}\cdot\mathrm{ACScore}

A.2 Multi-round Trajectory Refinement

Multinomial resampling uses softmax of negative score multiplied by a temperature parameter. Under this specific convention, increasing temperature sharpens selection and decreasing it promotes diversity—the reverse of formulas that divide logits by temperature. Subsequent partial noising and denoising explore around selected motions.

Equation
Prob=Softmax(−FinalScores⋅Temperature)\displaystyle\mathrm{Prob}=\mathrm{Softmax}(-\mathrm{FinalScores}\cdot\mathrm{Temperature})

A.3 Hyperparameters

The reported configuration uses twelve planning timesteps, five refinement rounds, temperature 1, initialization scale 0.001, and corner weight 0.2. Road and collision time weights decline from 1.0 to 0.85; proximity weights increase from 0.7 to 1.0. Settings are fixed across experiments except ablations and were chosen empirically without exhaustive search.

Hyperparameter Value
InitScale 1e-3
CornerWeight 0.2
H (planning horizon) 12 timesteps
DACStepWeight Quadratically decreasing from 1.0 to 0.85
ColStepWeight Quadratically decreasing from 1.0 to 0.85
ACStepWeight Linearly increasing from 0.7 to 1.0
DACWeight 5
CRWeight 10
ACWeight 1
Temperature 1.0
#Refinement Rounds 5
Table 3: Hyperparameters Used in Challenger. This table lists the empirical values of key hyperparameters used throughout our pipeline. These settings are selected based on general observations and remain fixed across all experiments (except for the ablation study), without task-specific tuning.

Appendix B Runtime Analysis

A typical six-view, roughly 240-frame scene takes 818.62 seconds on a server with two Xeon Platinum 8457C processors, four 48 GB NVIDIA L20 GPUs, and 512 GB RAM. Sampling takes 17.36 seconds, physics 211.36, scoring 194.72, and rendering 395.18. Trajectory sampling uses one GPU; rendering uses all four, while physics and scoring are CPU-bound. This is offline generation, about 13.6 minutes per scene on the stated hardware. Avoiding rendering each candidate is the principal cost-saving design; proposed GPU optimization of scoring is future work.

Original figure
Figure 8: Runtime Composition. Percentage of total time spent on each major component in the Challenger pipeline.
Component Runtime (s)
Traj. Sampling 17.36
Physics 211.36
Traj. Scoring 194.72
Rendering 395.18
Total 818.62
Table 4: Runtime Breakdown. Time spent on trajectory sampling, physics simulation (including the LQR controller and kinematic model), scoring, and multiview neural rendering.

Appendix C Video Quality Evaluation

The adapted VBench [68] protocol uses Grounded-SAM 2 with a car prompt, filters masks against projected three-dimensional annotations, then measures imaging quality and subject consistency for retained vehicles. This accommodates multiple changing subjects better than a single persistent-subject assumption. It remains annotation-assisted and vehicle-focused, so it does not independently validate the entire rendered environment.

Refer to caption
Figure 9: Video Quality Evaluation Pipeline. The video quality evaluation pipeline consists of two steps: (1) segmenting vehicles in the video frames using Grounded-SAM 2, and (2) computing the subject consistency metric and imaging quality metric from VBench.

Appendix D The Adv-nuSc Dataset

For each base scene, up to ten nearby background vehicles become candidate adversarial agents. Filtering removes collisions, departures from a 100-by-100-meter ego-centered area, and agents that never approach the ego vehicle. Remaining scenes are ranked by challenge before rendering, with manual inspection of a subset. The final dataset contains 156 scenes, 6,115 samples, 12,858 instances, and 225,085 annotations; it is a selected stress-test distribution rather than an estimate of natural event frequency.

Refer to caption
Figure 10: Example Scenes from the Adv-nuSc Dataset. We recommend the reader to have a look at our project page for dynamic videos.
Attribute Count
# instances 12,858
# ego poses 254,436
# scenes 156
# samples 6,115
# sample data 254,436
# annotations 225,085
Table 5: Overview of the Adv-nuSc Dataset. Summary statistics for our adversarial driving dataset generated using the Challenger framework. The dataset builds on the nuScenes validation split and includes rich 3D annotations, multi-camera frames, and diverse adversarial instances designed to evaluate the robustness of autonomous driving systems.

Appendix E Extending to Multiple Adversarial Vehicles

Multiple adversaries are added sequentially: generate one vehicle’s motion, feed the modified scene back into the pipeline, and choose another vehicle while treating earlier plans as fixed environment motion. Rendering occurs after all additions. The examples show richer simultaneous threats, but the method does not jointly model adversaries’ semantic coordination or solve a coupled multi-agent optimization.

Refer to caption
Figure 11: Extending to Multiple Adversarial Vehicles. The extension process is simple and modular. After generating a scenario with one adversarial vehicle and before rendering, we can simply feed the generated scene into Challenger again, and designate another vehicle as the adversarial vehicle.
Refer to caption
Figure 12: Example Scenes with Multiple Adversarial Vehicles. We recommend the reader to have a look at our project page for dynamic videos.

Appendix F Additional Qualitative Results

Additional UniAD and SparseDrive examples distinguish misforecasting from misplanning. In the first, other vehicles’ overtakes, cut-ins, blocking, or U-turns are predicted incorrectly. In the second, unsafe ego plans appear despite reasonably informative forecasts, including turns into occupied lanes or curbs. These manually selected cases illustrate failure types; they do not quantify their prevalence or establish which training change would eliminate them.

Refer to caption
(a)
Refer to caption
(b)
Refer to caption
(c)
Refer to caption
(d)
Figure 13: Misforecasting failure cases of UniAD [8].
Refer to caption
(a)
Refer to caption
(b)
Refer to caption
(c)
Refer to caption
(d)
Figure 14: Misforecasting failure cases of SparseDrive [10].
Refer to caption
(a)
Refer to caption
(b)
Refer to caption
(c)
Refer to caption
(d)
Figure 15: Misplanning failure cases of UniAD [8].
Refer to caption
(a)
Refer to caption
(b)
Refer to caption
(c)
Refer to caption
(d)
Figure 16: Misplanning failure cases of SparseDrive [10].
Concept demonstration

Open image ↗