Controllable risk scenario generation from human crash data for autonomous vehicle testing

2025-11-27

AI Summary

To address the lack of realism and controllability in risk-scenario generation for autonomous vehicle (AV) safety verification, this paper proposes a real-collision-data-driven method for synthesizing controllable risk scenarios. Methodologically, it introduces: (1) a novel decoupled risk latent-space modeling framework that disentangles normal traffic behaviors from critical-risk behaviors preceding collisions; (2) an optimization-driven smooth mode transition mechanism enabling long-horizon, tunable risk evolution; and (3) a risk-aware multi-agent co-optimization scheme guided by gradient-based refinement and calibrated via collision-data-driven distribution alignment to ensure behavioral fidelity. Experiments demonstrate that the method significantly outperforms baselines in scenario diversity, risk controllability, and physical plausibility. It generates high-fidelity, safety-critical scenarios with precise risk targeting while preserving naturalistic everyday traffic dynamics—thereby substantially improving coverage and efficiency in AV robustness assessment.

Technology Category

• Search and Optimization: Sampling/Simulation-based Search
• Humans and AI: Human-Aware Planning and Behavior Prediction
• Planning, Routing, and Scheduling: Model-Based Reasoning

Application Category

• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Search and Retrieval-Augmented AI: Retrieval-Augmented Generation (RAG) and multi-modal RAG

Abstract

Ensuring the safety of autonomous vehicles (AV) requires rigorous testing under both everyday driving and rare, safety-critical conditions. A key challenge lies in simulating environment agents, including background vehicles (BVs) and vulnerable road users (VRUs), that behave realistically in nominal traffic while also exhibiting risk-prone behaviors consistent with real-world accidents. We introduce Controllable Risk Agent Generation (CRAG), a framework designed to unify the modeling of dominant nominal behaviors and rare safety-critical behaviors. CRAG constructs a structured latent space that disentangles normal and risk-related behaviors, enabling efficient use of limited crash data. By combining risk-aware latent representations with optimization-based mode-transition mechanisms, the framework allows agents to shift smoothly and plausibly from safe to risk states over extended horizons, while maintaining high fidelity in both regimes. Extensive experiments show that CRAG improves diversity compared to existing baselines, while also enabling controllable generation of risk scenarios for targeted and efficient evaluation of AV robustness.

Problem

• Generates realistic risk scenarios from crash data for AV testing
• Models both normal and risk-prone behaviors in traffic agents
• Enables controllable risk generation for targeted AV robustness evaluation

Innovation

• Controllable Risk Agent Generation framework
• Structured latent space disentangles normal and risk behaviors
• Risk-aware latent representations with optimization-based mode transitions