It is the second time deepsense.ai researchers have presented at the NeurIPS conference. In the most recent installment of the world’s most closely watched AI conference, Piotr Miłoś described CRTS research during the Machine Learning for Autonomous Driving workshop. The presentation continued deepsense.ai’s fruitful cooperation with Volkswagen on training autonomous vehicles in a simulated environment.
The R&D project was conducted by a team of researchers including deepsense.ai’s Miłoś and Michał Martyniak, as well as Błażej Osiński, Adam Jakubowski, Paweł Zięcina, Christopher Galias, Henryk Michalewski together with Volkswagen researchers Silviu Homoceanu and Antonina Breuer.
“During this stage of the project we developed a port of real-life traffic footage into the CARL simulator, which we called CARLA Realistic Traffic Scenarios (CRTS). The aim was to provide a training and evaluation ground environment for obtaining better driving policies. We believed that an agent developed within the CRTS scenarios would behave more naturally when deployed, due to the simple fact that it has had a chance to encounter more realistic situations while training,”
explained Miłoś, who in addition to working as a researcher at deepsense.ai is a professor at the Polish Academy of Sciences.
The collaborative research has led to a number of noteworthy breakthroughs. The artificial agent obtained with reinforcement learning has become “more dynamic” than the reference human driver. That it drives faster is one example of how. Being able to use simulation for such experiments offers crucial advantages, including being less expensive and less time-consuming. Over the combined experiments, the team generated over 100 years of simulated driving experience.
The observations that resulted from the experiment bring valuable benefits both for scientific and commercial projects focused on real-life developments of autonomous cars. Details of the research are described in the team’s paper as well as on the project’s website.