A Local Agentic AI Framework to Generate Edge Case Scenarios for Safety Assurance

TL;DR

This paper proposes a local agentic AI framework for generating edge case scenarios to ensure safety assurance in autonomous systems. The framework uses language models to generate and refine scenarios, which are then simulated and evaluated. This approach enables adaptive and automated edge case generation.

Why It Matters

The proposed framework has practical significance as it can help identify hazardous failures and increase system reliability and robustness. It also ensures data privacy and security, making it a valuable tool for safety-critical applications. By automating edge case generation, the framework can reduce the workload and costs associated with traditional human-in-the-loop approaches.

Key Findings

• adaptive scenario generation using local language models
• refinement of scenarios using domain-specific criteria
• identification of functional gaps and triggering conditions through post-hoc clustering