Autonomous driving systems rely upon sensors and embedded software for localization, perception, motion planning and execution. They can only be released to the public after developers have demonstrated their ability to achieve high levels of safety.
Today’s hands-off autonomous driving systems are largely built with deep learning algorithms that can be trained to make the right decision for nearly every driving situation. These systems, however, lack the detailed requirements and architecture that have been used up until now to validate safety-critical software, such as the kind that control commercial airliners. Road testing is clearly an essential part of the development process, but billions of miles of road testing would be required to validate the safety of autonomous driving systems and software. Simulation serves the need to make verification and validation of the operation of autonomous vehicles a practical effort.