Evaluating Physical Consistency and Plausibility in Generative Scenario Models for Autonomous Driving

2026-10-01

AI Summary

This study addresses the lack of physical consistency verification in generative autonomous driving scenario models, which impedes their compliance with safety certification requirements. To overcome this limitation, this work proposes a pioneering five-layer evaluation protocol. By integrating variational autoencoders (VAEs), kinematic alignment, and activation analysis techniques, the proposed method comprehensively assesses the alignment between latent representations and vehicle dynamics, spanning from internal representations to output dynamic constraints. Experimental results demonstrate that this protocol effectively uncovers latent physical inconsistencies imperceptible to conventional visual metrics. Furthermore, its generalizability is validated across diverse generative architectures. Ultimately, this research establishes a novel paradigm for the safety evaluation of generative models in autonomous driving.

Abstract

Generative AI models are increasingly used for scenario generation in autonomous driving. While they can generate realistic-looking scenarios, they often provide limited transparency into learned representations and consistency with real-world vehicle dynamics. This lack of formal assurance limits their use in safety-critical validation and certification workflows. To address this aspect, we introduce a layered evaluation protocol that complements existing methods by assessing models across five layers. The first four layers inspect internal representations and network layers through kinematic alignment, statistical baseline comparison, latent controllability, and activation analysis. The fifth layer evaluates model outputs against vehicle dynamics constraints such as lateral jerk thresholds. We demonstrate the protocol on a Variational Autoencoder (VAE)-based scenario generator. Although standard output-level metrics and visualizations suggest that the generated scenarios are realistic, our protocol provides deeper insight into the extent to which the model's latent space aligns with kinematic features and whether visually plausible trajectories satisfy vehicle-dynamics constraints. We further apply the protocol to additional generative models, demonstrating its applicability beyond the VAE architecture.

Problem

• Generative AI
• Autonomous Driving
• Physical Consistency
• Scenario Generation
• Safety Validation

Innovation

• Layered Evaluation Protocol
• Physical Consistency
• Generative Scenario Models
• Vehicle Dynamics Constraints
• Latent Controllability