Testing agents and controlling vehicles from simulation to real applications

Abstract

Validating an autonomous driving agent is a complex multi-step process that requires close collaboration between the software developer the vehicle provider. The purpose of this talk, is to introduce the audience to a common workflow among industries and research institutes to efficiently deploy custom software (both on the vehicle control side and on the higher-level autonomous driving side) also adopted in the H2020 Dreams4Cars research project. The first step is the setup of a Model-in-the-Loop (MIL) environment. In MIL, all the building blocks of the software architecture run in a simulated environment thus neglecting real-time constraints, sensor noise and the inherent stochastic nature of a real road. Once the functioning of the software has been thoroughly tested in such an environment, the next step is to move the newly designed modules to the specific hardware for the vehicle implementation to fulfill the Software-in-the-Loop (SIL) environment. Before the actual vehicle deployment, a third step is the set-up of a real-time machine capable of generating CAN messages as found in the car, also known as Hardware-in-the-Loop (HIL) test bench. The presentation will also propose two simulation environments and their usage in the context of the described workflow. Finally, a new neural-network based approach to the vehicle control which tries to overcome traditional model-based controller will be presented.