Regulators need to have confidence in how different developers use AI, which is why we need independent safety validation research. DriveSafeAI will address this through four technical work packages focused on AI safety assurance, covering methodologies from scenario generation to model evaluation.
Evidence for a generalizable AI safety framework for self-driving, aligned with internationally recognized ODD-based safety assurance methods. This includes evidence of the completeness of data for scenario training and testing.
We test various approaches to scenario generation and publish selected scenarios on Safety Pool™. This supports industry efforts to build a shared scenario library for training and testing.
The industry currently lacks a detailed methodology to prove the trustworthiness of virtual test environments and methods for correlating test results from simulation with the real world. We will develop evidence to inform simulation validation guidelines by comparing model performance in real-world and simulated environments.
We will conduct a safety analysis based on the reference architecture (ISO/TS 5469) to develop a design rationale for a redundant safety architecture in AI-based systems and share safety recommendations for the industry.
