STRELGen generates safety-critical driving scenarios with spatio-temporal logic

A team led by Lorenzo Bonin, Francesco Giacomarra, Luca Bortolussi, Jyotirmoy V. Deshmukh, and Francesca Cairoli has published STRELGen, a scalable framework that rethinks autonomous vehicle safety validation. Traditional approaches rely on exposing AD systems to millions of real-world traffic scenes—a brute-force method that is both expensive and statistically poor at capturing the rare, safety-critical edge cases needed to ensure robustness. STRELGen tackles this by synergistically combining a multi-agent trajectory-generation diffusion model with Spatio-Temporal Logic (STREL) specifications. STREL encodes complex safety and realism properties in a highly interpretable formalism, and crucially, its monitoring satisfaction levels are differentiable.

At inference time, STRELGen optimizes directly over the diffusion model's latent space to maximize satisfaction of the STREL formula, yielding efficient generation of highly plausible yet safety-critical multi-agent scenarios that remain within the learned data distribution. This gradient-based search enables targeted creation of dangerous but realistic scenarios—such as near-misses or sudden pedestrian crossings—that traditional methods would require enormous datasets to encounter. STRELGen thus provides a flexible, interpretable, and powerful tool for stress-testing autonomous driving systems, moving beyond the limitations of brute-force data collection. The paper is available on arXiv (2605.19038) and represents a significant advance in neuro-symbolic methods for safety validation.