An Analysis of Software Latency for a High-Speed Autonomous Race Car—A Case Study in the Indy Autonomous Challenge

Autonomous driving faces the difficulty of securing the lowest possible software execution times to allow a safe and reliable application. One critical variable for autonomous vehicles is the latency from the detection of obstacles to the final actuation response of the vehicle, especially in the case of high-speed driving. A prerequisite for autonomy software is that it enables low execution times to achieve superhuman reaction times. This article presents an in-depth analysis of a full self-driving software stack for autonomous racing. A modular software stack especially developed for high-speed autonomous driving is used and the latency of the software is analyzed in four main autonomy modules: perception, prediction, planning, and control. With the help of a trace point measurement method, it is possible to investigate the end-to-end latency and runtimes of the individual modules. This analysis is conducted for different scenarios (high-speed runs, object avoidance, and multi-vehicle racing) and is based on the two sensor pipelines. With this evaluation, insights are provided that highlight the importance of software execution times for several parts of the software architecture. Real data from an actual vehicle and simulation data show that both complex and linear relationships exist, from which optimization potential for future software stacks can be derived.