Learnable Operational Design Condition monitor for failure prediction in autonomous driving

Abstract

Autonomous vehicle accidents frequently originate from violations of Operational Design Conditions (ODC)—the predefined operational limits of vehicle states, environmental factors, and driver capabilities. ODC violation indicates a high probability of impending functional failure under current operational scenarios, where failure denotes the system’s inability to achieve designated performance thresholds or maintain safety-critical constraints. Therefore, designing and continuously monitoring the boundary states of ODC during system operation constitutes a critical imperative to prevent system failures, ensure operational safety, and mitigate autonomous vehicle accidents. However, prevailing ODC monitoring methods primarily rely on first-order logic checklists, failing to capture emergent risks from parameter interactions. Therefore, this paper establishes an end-to-end methodological framework for constructing a learnable ODC monitor termed ODCNet to realize failure prediction. The architecture first projects operational states into unified latent representations, then derives probabilistic boundary estimates through neural inference, and finally calibrates residual errors via hybrid Gaussian Process regression. Additionally, an adaptive active learning mechanism continuously refines boundary precision through targeted testing of high-uncertainty scenarios. The validation through intersection, lane-keeping, and vehicle detection case studies demonstrates the failure prediction performance of the ODC monitor that precedes accidents.