I-Corps: Translation potential of a verification software module that evaluates end-to-end driving models for autonomous vehicles

This I-Corps project is based on the development of a formal verification framework that mathematically evaluates the safety of artificial intelligence models for autonomous vehicles before they are deployed. Current validation methods for autonomous systems rely on running billions of simulated scenarios to find failures, which cannot account for the effectively infinite number of unpredictable cases that can occur in the real world. To address this massive challenge, this technology provides a proactive software utility designed to audit the internal structure of an artificial intelligence model and predict unstable operating regions before any simulations are run. By shifting from reactive testing to proactive verification, this technology may significantly accelerate the safe and certifiable deployment of advanced automated systems, ultimately helping to reduce traffic crashes and fatalities. In addition, this technology has potential application in other areas, offering a way to ensure the bounded, safe operation of general robotics, manufacturing processes, and advanced generative artificial intelligence applications. This I-Corps project utilizes experiential learning coupled with first-hand investigation of the industry ecosystem to assess the translation potential of a formal verification software framework for autonomous vehicles. The goal of this technology is to bridge the gap between advanced mathematical verification and applied industrial software engineering by investigating the core friction points validation teams face. This technology is based on semidefinite programming relaxations combined with Topological Data Analysis to compute certified safety bounds for end-to-end neural networks. The technical results show that this deterministic method guarantees mathematically strict performance outcomes. This method is different from existing simulation-based solutions in its ability to mathematically map a neural network's latent manifold coverage to proactively flag unstable operating regions without relying on statistical sampling. This may transform theoretical mathematical bounds into a robust, verification-aware architecture that provides a foundation for the scalable safety certification of complex artificial intelligence systems. Users may deploy this technology as an on-premise plug-in that provides a ranked instability report, enabling engineers to prioritize testing and drastically reduce computing costs. This award reflects NSF's statutory mission and has been deemed worthy of support through evaluation using the Foundation's intellectual merit and broader impacts review criteria.