Enhancing the Safety of Autonomous Driving Systems via AoI-Optimized Task Scheduling

An Autonomous Driving System (ADS) uses various sensors and deep learning to improve navigation and control tasks. Maintaining road safety requires that these tasks are seamlessly synchronized and consistently utilize the most recent sensing data. This synchronization poses challenges due to 1) various sensing periods across different sensors, 2) the interdependency of tasks, and 3) constraints on computational resources. Our research pioneers the use of the Age of Information (AoI) to measure task scheduling performance within ADS. With theoretical analysis, we disclose that optimizing AoI simultaneously minimizes response time and maximizes throughput. We then offer a formal definition of the AoI-centric task scheduling problem. Given the NP-hardness of this problem, we design a 4-approximation algorithm. To enhance the practicality of our solutions, we propose an extended formulation optimizing AoI-centric scheduling over a specified cycle and correspondingly develop a reinforcement learning-based approach. Experimental simulations, benchmarked against the Apollo driving system, demonstrate that our AoI-optimized task scheduling outperforms Apollo's scheduling mechanisms in terms of AoI, throughput, and worst-case response time. Notably, our proposed solution operating on four cores yields a maximum AoI lower than Apollo's schedulers running on eight cores.