Risk-Controllable Multi-View Diffusion for Driving Scenario Generation

Lin, Hongyi; Shi, Wenxiu; Huang, Heye; Zhuang, Dingyi; Zhang, Song; Liu, Yang; Qu, Xiaobo; Zhao, Jinhua · 2026 · arXiv (Cornell University)

DOI: 10.48550/arxiv.2603.11534

archive: archived pipeline: cataloged verified

Get this paper ↗ (DOI — opens at the source; we link to it, we don't host it)

Summary

This paper addresses the challenge of generating safety-critical, long-tail driving scenarios for autonomous vehicle evaluation. Existing methods struggle to explicitly control risk levels, often treating risk as an after-the-fact label, and frequently fail to maintain geometric consistency across multiple camera views. The authors propose RiskMV-DPO, a pipeline that reformulates driving risk as a proactive, time-resolved control signal to autonomously synthesize diverse, high-stakes dynamic trajectories. The method operates in two stages. First, a risk-control module generates physically-grounded motion trajectories and 3D bounding boxes conditioned on a user-specified target risk level. This module uses a field-inspired risk computation that accounts for relative velocity, agent type, and interaction direction, and employs feature-wise linear modulation to guide a motion generator. Second, a multi-view diffusion model renders temporally coherent video sequences conditioned on these motion controls. To ensure geometric fidelity, the authors introduce a geometry-appearance alignment module that injects compact 3D priors from a VGGT backbone into the diffusion process. Additionally, they develop a region-aware direct preference optimization (RA-DPO) strategy with motion-aware masking. This approach focuses learning on dynamic, motion-dominant regions by fusing temporal motion cues with geometrically consistent masks derived from agent trajectories, thereby improving spatial-temporal coherence. Experiments on the nuScenes dataset demonstrate that RiskMV-DPO outperforms state-of-the-art baselines such as DriveDreamer-2 and MagicDriveV2. The model achieves a Fréchet Inception Distance (FID) of 15.70, indicating superior visual quality, and improves the mean Average Precision (mAP) for 3D detection from 18.17 to 30.50, reflecting enhanced geometric realism. Ablation studies confirm that the geometry-appearance alignment and motion-aware masking significantly contribute to these gains, improving multi-view structural similarity and depth accuracy. The significance of this work lies in shifting world models from passive environment prediction to proactive, risk-controllable synthesis. By enabling the generation of specific, high-risk scenarios with high geometric fidelity, RiskMV-DPO provides a scalable toolchain for stress-testing and safety-oriented development of embodied intelligence systems. This approach allows for the efficient exploration of rare, hazardous interactions that are underrepresented in real-world data, addressing a critical bottleneck in autonomous driving validation.

Key finding

The proposed RiskMV-DPO framework enables the generation of diverse, risk-controllable driving scenarios that improve 3D detection mean average precision from 18.17 to 30.50 and reduce Fréchet Inception Distance to 15.70 on the nuScenes dataset.

Methodology

simulation_modeling

Provenance

The full processing record for this entry. Every stage of this paper's journey through the pipeline is logged — what ran, with which tool and model, how many attempts it took, and when it last completed. Discovered via author_sweep_intake on 2026-05-28.

Summary generated by qwen3.6-27b-prismaquant on 2026-06-10; verification: verified.

Topics

Ranked by relevance to this paper. Hover a topic for its definition.

• situational awareness

Information type

What kind of knowledge this paper contributes, grouped by family — independent of topic (what it is about) and method (how it was studied).

• Theoretical Contribution: computational model