Bootstrapped Neural Models for Predicting Self-Driving Vehicle Collisions With Quantified Confidence: Offline and Online Applications

Highly automated vehicles are complex systems, and ensuring their safe operation within their Operational Design Domain (ODD) presents significant challenges. Diagnosing failure modes and updating these systems are even more demanding tasks. This paper introduces a method to assist with assessing, diagnosing, and updating these systems by developing a stochastic model that predicts safety outcomes (collision, near-miss, or safe state) with quantified uncertainty in any parametrized scenario. The approach uses bootstrapping aggregation to create an ensemble of predictive models, leveraging fully connected feed-forward neural networks. These networks are designed with a flexible number of trainable parameters and hidden layers, requiring minimal computational resources. The model is trained on a small set of examples obtained through direct simulations that randomly sample the parametric scenario, bypassing the traditional test matrix definition. Once trained, the bootstrapped model serves as an identity card for the system under test, allowing continuous performance evaluation across the parametric scenario. The paper demonstrates applications, including safety assessment, failure mode identification, and developing a safe speed recommendation function. The model's compact size ensures rapid execution, facilitating extensive post-analysis for safety argumentation and diagnosis and real-time online use to extend the system's abilities.