Sim City Created for Self-Driving Cars
A virtual city simulation trains driving A.I.s to recognize and deal with obstacles, pedestrians and unpredictability.
It's like a video game for an A.I.
In this case, the artificial intelligence is software that would ultimately be used to control a self-driving car. And the video game is a simulation of a city rife with unpredictable driving conditions, just like real life.
The simulation, called Synthia, was developed by scientists at the Computer Vision Center in Barcelona. It has all of the features you'd expect to see in a bustling city, including other vehicles, pedestrians, cyclists, double-parked buses. But even more than that, it introduces variables such rain and snow.
A good snow cover can blur the lines between a road and a sidewalk and so building an A.I. that's able to recognize the difference is a good step in the direction of smarter self-driving cars.
"A.I.'s are becoming very good at recognizing objects such as pedestrians or vehicles," team leader German Ros told Gizmag. "However, the boundaries of sidewalks and the recognition of traffic lights are still very challenging.
The researchers added a virtual car to the simulated environment complete with cameras designed to shoot video and still images from the car's perspective.
WATCH VIDEO: How Safe Are Self-Driving Cars?
As the car drove through the environment collecting video and images, software identified the objects and conditions and annotated them.
Until now, images used to train artificial intelligence programs are collected under real world conditions and annotated manually.
Accurately annotating video and images is tedious work. What's more, infrequent events, such as having a bus pull out in front of you or having a cyclist veer into your lane unexpectedly, are not always captured in real-world video or images and so artificial intelligence programs don't get good training about these conditions.
After collecting more than 213,000 images and video sequences in the virtual world, Ros and his team analyzed whether they really improved an A.I.'s ability to recognize similar events in the real world. Turns out they did. The A.I.'s success rate went from about 45 to around 55 percent.
To improve the software even more, Ros and his team are releasing all of the data produced by Synthia for non-commercial use. Hopefully other research teams will build on it.
That's good news for future self-driving vehicles, because safety is no game.