An End-to-End Optimization Framework for Autonomous Driving Software

Given the increasing complexity of autonomous driving, it becomes more difficult to test driving functions and to optimize algorithm parameters. One major challenge is that many parameters and software components influence each other, so even small changes in parameters can lead to a high sensitivity in vehicle performance. Many approaches involve real-world and simulation-based testing of predefined scenarios, which is expensive and time-consuming, and manually determining of reliable software parameters is not possible in many applications because parameter variation is non-intuitive. Misconfigurations of the software parameters are detected too late. For that reason, reliable and automated software testing and optimization is an essential component for autonomous driving in the future. This paper presents an end-to-end optimization framework for automatically tuning and optimizing individual parameters for a full-stack autonomous driving software. We will demonstrate our method for optimizing the parameters in a non-deterministic simulation environment by using gradient-free optimization methods. The simulative method we are presenting was applied and deployed at the Indy Autonomous Challenge. This method offers the opinion of building a remote tool chain that efficiently supports testing and optimization under dynamic requirements during the autonomous driving software development process.