Machine Learning based Testing and Validation of Automated Vehicles
Date:
Abstract:
Current research and development aims to increase the amount of automation in vehicles, to ultimately achieve fully automated driving. In order to complete the shift of the driving task from the human driver to the machine, it needs to be ensured that the automation system can deal with all the critical scenarios that can potentially arise. A safety argument needs to be built through validation and testing, which proves the maturity of automated vehicles for introduction into public use. Basing such a safety argument solely on the covered distance during testing alone is however infeasible, as an estimated 11 billion km of driving are necessary to prove that an automated vehicle is safer than the average human driver. A scenario-based approach, where it is proven that a sufficient number and type of scenarios have been covered in the validation, is much more practical.
At the Centre for Connected and Automated Automotive Research (CCAAR), a collaboration between Coventry University and HORIBA-MIRA, we are exploring the problem of validating automated vehicles. Slight changes in the interaction of the system with other traffic participants and the environment lead to an infinite number of possible scenarios. We apply supervised machine learning alongside scenario simulation to explore this scenario space and find the critical corner case scenarios. These critical corner case scenarios flag up design flaws in the automated system and can guide testing activities in subsequent higher fidelity tests such as Hardware-in-the-Loop and proving ground trials.