Virtual environment for training autonomous vehicles

Abstract

Driver assistance and semi-autonomous features are regularly added to commercial vehicles with two key stakes: collecting data for training self-driving algorithms, and using these vehicles as testbeds for these algorithms. Due to the nature of algorithms used in autonomous vehicles, their behavior in unknown situation is not fully predictable. This calls for extensive testing. In this paper, we propose to use a virtual environment for both testing algorithms for autonomous vehicles and acquiring simulated data for their training. The benefit of this environment is to able to train algorithms with realistic simulated sensor data before their deployment in real life. To this end, the proposed virtual environment has the capacity to generate similar data than real sensors (e.g. cameras, LiDar,..). After reviewing state-of-the-art techniques and datasets available for the automotive industry, we identify that dynamic data generated on-demand is needed to improve the current results in training autonomous vehicles. Our proposition describes the benefits a virtual environment brings in improving the development, quality and confidence in the algorithms.