An Autonomous Driving System (ADS) uses a plethora of sensors and many deep learning based tasks to aid its perception, prediction, motion planning, and vehicle control. In recent years, various models and algorithms have been developed for individual tasks to achieve shorter execution time and higher accuracy. When such tasks are put into a pipeline, the overall performance of the system depends on not only the performance of individual tasks, but also the scheduling of the tasks on multi-core processors. In a broad sense, scheduling of autonomous driving tasks falls into the category of task scheduling for real-time stream processing. The performance metrics commonly considered in this area are response time and throughput. While in autonomous driving, driving safety is the first priority, which imposes a few new challenges and makes task scheduling in autonomous driving different from that for conventional real-time stream processing. To ensure driving safety, tasks should be synchronized and use the latest sensing data, which is challenging since 1) different sensors have different sensing periods, 2) the tasks are inter-dependent, 3) the computing resource is limited.

This thesis derives a new task scheduling policy, AoI-centric task scheduling, from the perspective of driving safety by first introducing Age of Information (AoI) as the performance metric for task scheduling in an ADS, which is defined as the time difference between the time when the control task generates control commands and the minimum raw data timestamp in the inputs of the control task.

First, we show that the objective of minimizing the maximum AoI is equivalent to jointly maximizing the throughput and minimizing the response time, where throughput and response time are often separately considered in conventional task scheduling. We formally formulate the AoI-centric task scheduling problem as an ILP problem, referred to as ILP-AoI, where the optimal solution in a small scheduling horizon can serve as the lower bound of the large scheduling horizon. To derive practical scheduling solutions, we then extend the formulation and formulate the optimal AoI-centric periodic scheduling problem with a given cycle, referred to as ILP-AoI-II.

Second, we develop a novel reinforcement learning (RL) based scheduler to efficiently solve the ILP-AoI-II problem. The RL solution fully leverages the properties derived from ILP-AoI-II formulation and the domain knowledge of ADSs to solve the hard combinatorial optimization problem. We student the performance of the proposed RL-based scheduling with two event-driven scheduling approaches in the Apollo system and the experiment results show that the maximum AoI in the proposed scheduling solution with 4 cores is lower than that in Apollo’s schedulers with 8 cores.

The proposed AoI-centric task scheduling in this thesis will make contributions not only to autonomous driving, but also other real-time applications with similar natures.