M-DRTA: A Distributed Runtime Monitoring and Assurance Framework for Multi-vehicle Behavior Planning

The machine learning-based decision algorithms commonly used in autonomous driving have led to the Safety of the Intended Functionality(SOTIF) issues due to their potential lack of functionality. To address these limitations, we propose M-DRTA, a distributed runtime assurance framework based on machine learning, which can provide safety assurance for multi-autonomous vehicles by making functional improvements to narrow the vehicle safety zone. We secure the entire multi-vehicle driving system by maintaining safety in local vehicles. In the M-DRTA, an independent runtime assurance framework is provided for each autonomous vehicle through redundant functional modules that include a deep neural network-based advanced controller, a recoverable safety controller, and a monitoring and assurance module. Monitor SOTIF risks by identifying trigger conditions, under-function status, etc., and provide safety through permission handover without unduly sacrificing performance. We tested and evaluated M-DRTA for the different number of vehicle states in a driving task. The experimental results show that M-DRTA can strike a proper balance between safety and efficiency compared to the baseline approach.