Researchers propose neuro-symbolic safety guard for autonomous driving

Researchers Simón Patiño Idarraga, Erick Silva, Rehana Yasmin, and Ali Shoker have proposed a neuro-symbolic safety guard to address critical safety gaps in end-to-end autonomous driving systems. These systems, while achieving high average performance, often fail to adhere to fundamental traffic rules due to their reliance on statistical pattern recognition rather than physical safety constraints. The proposed solution introduces a lightweight module that attaches to the final command interface of an already-trained agent, acting as a last-line safety check before commands are executed.

The neuro-symbolic safety guard enforces explicit safety rules in real-time, replacing unsafe commands with the nearest safe alternative only when necessary. This approach ensures traceability, as each intervention is directly linked to the specific rule that triggered it. In evaluations using the state-of-the-art TransFuser v6 model on long-tail benchmarks Fail2Drive and Bench2Drive, the safety guard improved the success rate by 15% and reduced safety-critical collisions by up to 53%, all while preserving the original driving score. Crucially, the guard requires no retraining and introduces no additional learned components.