Realistic Simulation of Vehicles for safer, more robust and cheaper development of autonomous vehicles
|Funding from Vinnova||SEK 1 772 000|
|Project duration||November 2017 - December 2019|
|Venture||Machine Learning - FFI|
|Call||Machine Learning - FFI - 2017-06-13|
Purpose and goal
One of the largest challenges in developing autonomous driving systems is access to data. We want to develop two simulation tools in this project, one of which is based on machine learning and builds on the first one. The tool will learn, by means of unsupervised machine learning, to translate synthetic data from simulation to more realistic data. The tool should be used to train intelligent behaviour by agents, through reinforcement learning, as well as generate huge amounts of training data for supervised machine learning.
Expected results and effects
We expect to see an improvement in the performance of the trained systems when using our simulated data. Furthermore, we expect performance to be further enhanced when we use the refined synthetic data. By seeing how much of the annotated data we can ignore and still achieve the same performance, through the combination of synthetic data, we can quantify cost savings.
Planned approach and implementation
We will equip vehicles with suitable sensors. Data will then be logged from this vehicle in different environments and conditions. A subset of this data is then annotated. At the same time, we develop the simulation software to support simulations with these sensors. Then we develop the machine learning system that refines the synthetic data. Finally, the project will be evaluated through real-life experiments on the AstaZero test track.