Grammar-Constrained Refinement of Safety Rules for Autonomous Systems
Researchers propose a framework for refining safety operational rules in cyber-physical systems (CPS) using language-in-the-loop and counterfactual reasoning. The method addresses the challenge of maintaining syntactic correctness under domain-specific grammars while avoiding overfitting to observed outcomes. Applied to an autonomous driving control system, the approach resolved inconsistencies between operational rules and observed system behavior during simulation-based verification. The work highlights risks of semantically unjustified refinements that could compromise safety.
Key facts
- Safety specifications in CPS capture operational conditions for safe operation.
- Operational rules must be refined as environments evolve to maintain consistency with observed behavior.
- Revising inconsistent rules requires syntactic correctness under a domain-specific grammar.
- Language-in-the-loop refinement can produce semantically unjustified refinements that overfit.
- The framework combines counterfactual reasoning with a grammar-constrained refinement loop.
- Applied to an autonomous driving control system.
- The approach successfully resolved inconsistencies.
- arXiv paper number: 2604.23523.
Entities
Institutions
- arXiv