A group of researchers of the English Wayve startup has developed a method of network application of deep learning to Autonomous driving. Not so long ago representatives of the company demonstrated how their technology works, and has offered to bring a real car on a real road and teach him… self driving for 20 minutes.
The fact is that most self-driving cars uses a set of cameras and sensors, as well as tools for mapping and computer programming. But this approach, in their opinion, rests on the ceiling. Autonomous vehicles are programmed by companies such as Google, have reached the point where they are good, but not good enough for widespread use. This is due to the fact that the car is not smart enough to cope with lots of conditions on an ordinary road. According to scientists, a car needs more than a clever computer, but not special sensors or programming.
The Wayve team believes that a more sensible approach is to use learning algorithms with reinforcements, something that will involve DeepMind is to let the computer learn like humans, in practice. Algorithms of reinforcement learning are the basis of deep learning networks they are studying, repeating and repeating the task, each time improving your score. In the case of Autonomous vehicle control, this would mean driving the car until then, until he learns to do it correctly.
To demonstrate how well it could work like this approach, Wayve team Renault Twizy is equipped with one camera, control gas, brake and steering wheel, and connected it to the GPU and the computer algorithms of reinforcement learning. The computer “said” that the best result would be a car moving on the road without leaving the road. The longer he does, the better. Then they added the driver and put the car on a country road. The computer has taught a car to drive, not skidding off the road in 20 minutes.
A car with deep learning self-learned driving in 20 minutes
Ilya Hel