Machine Learning Based Methods for Virtual Validation of Autonomous Driving

Licentiate thesis, 2021

During the last decade, automotive manufacturers have introduced increasingly capable driving automation functions in consumer vehicles. As the functionality becomes more advanced, the task of driving moves from the human to the car. Hence, making sure that autonomous driving (AD) functions are reliable and safe is of high importance. Often, increased levels of automation result in more complex safety validation procedures, that may be both expensive, time consuming, and dangerous to perform. One way to address these problems is to move parts of the validation to the virtual domain.

In this thesis, we investigate methods for validating AD functionality in virtual simulation environments, using methods from machine learning and statistics. The main focus is on how to make virtual simulations resemble real-world conditions as closely as possible. We tackle this with an approach based on sensor error modeling. Specifically, we develop a statistical sensor error model that can be used to make ideal object measurements from simulations resemble measurements obtained from the perception system of a real-world vehicle. The model, which is based on autoregressive recurrent mixture density networks, was trained on sensor error data collected on European roads.

The second part considers system falsification using reinforcement learning (RL); a flexible framework for validation of system safety, which naturally allows for the integration of, e.g., sensor error models. We compare results of system falsification using RL to an exact approach based on reachability analysis.

With this thesis, we take steps towards more realistic statistical sensor error models for virtual simulation environments. We also demonstrate that approximate methods based on reinforcement learning may serve as an alternative to reachability analysis for validation of high-dimensional systems. Finally, we connect the RL falsification application to sensor error modeling as a possible direction for future research.