Neuro-Symbolic Safety Guards for End-to-End Autonomous Driving
A recent study presents a neuro-symbolic safety mechanism designed for end-to-end autonomous driving systems, tackling the problem of agents breaching fundamental traffic regulations despite achieving high average performance. This safety guard is a compact module that integrates with the final command interface of a pre-trained agent, validating each command against specific safety protocols and substituting it with the closest safe option when required. This method is directly actionable and linked to the corresponding rule, eliminating the need for retraining or additional learned components. The research assessed the system using long-tail benchmarks Fail2Drive and Bench2Drive, employing the advanced TransFuser v6 (TFv6) as a case study. The paper is accessible on arXiv with the identifier 2608.11451, noted as a cross-type submission. The proposed method seeks to enhance decision-making transparency and explicitly enforce safety constraints, differing from purely statistical learning approaches.
Key facts
- The paper introduces a neuro-symbolic safety guard for end-to-end autonomous driving.
- The guard attaches to the final command interface of an already-trained agent.
- It checks commands against explicit safety rules and replaces unsafe commands with the nearest safe alternative.
- Each intervention is traceable to the rule that triggered it.
- The guard requires no retraining and adds no learned component.
- Evaluated on Fail2Drive and Bench2Drive benchmarks using TransFuser v6 (TFv6).
- The paper is available on arXiv with identifier 2608.11451.
- The approach addresses the structural issue of agents learning statistical patterns rather than physical safety conditions.
Entities
Institutions
- arXiv