At California-based Ford Research and Innovation Center, scientists are turning to video games to improve the company’s autonomous vehicle software. In the Palo Alto center, you will find researchers intently studying a large TV screen, showcasing a robotic humanoid racing a a four-wheel car. Running much faster than a real human, the humanoid-like figure abruptly cuts in front of the car a couple of times, forcing it to back into the sidewalk to avoid an accident.
The research is part of the scientists’ attempt to teach Ford’s autonomous vehicle software how to better handle real-life situations, involving actual cars zooming down the highway. The team is currently using game development tools to create special virtual environments and action sequences, which are in turn fed into advanced machine learning systems. This approach significantly increases the software’s speed of learning, taking only 20 minutes for something that would have required the researchers 10 days to achieve.
To enhance the ease of research work, the scientists have also developed a new version of the simulation software that is compatible with large gaming systems as well as mobile phones. In addition to learning how to avoid hitting pedestrians, the virtual environment also teaches the software how to identify traffic lanes when there are no markings on the road. Speaking about the project, Tory Smith, a member of the research team, said:
We’re trying to frustrate our system… We can hand the mobile version to random people on the street, [as they] can think of weird things to do to challenge the vehicle.
According to CEO Mark Fields, the company will begin testing its self-driving car prototypes in California, as early as 2016.
Via: IEEE Spectrum