artificial intelligence that learns to drive in 20 minutes

Autonomous cars have a battery of sensors, which tell you how to drive, with a long list of carefully designed rules and thoughts. What if we let a car learn to follow its path? And if, as and when trial and error, the car managed to learn and build his driving? This is what Wayve proposes.

" In only 15 to 20 minutes, it is possible to teach a car to follow a line and react, starting from a blank sheet, simply by the training she carries out and the feedback that her pilot gives her »explain the creators of Wayve

Learning to drive with reinforcement learning

Wayve is an English start-up created by researchers at the University of Cambridge, which uses reinforcement learning. No 3D mapping of the road, no rules and no upstream development. One of the reference algorithms, DDPG, has been implemented and improved, so that the car learns to "hold its line". A single visual source processed by an onboard graphics processor allows the Renault Twizy used by the researchers to grope and learn.

The car winds and gets out of line. The driver corrects. The car makes a mistake. He corrects again. Self learns, and understands that staying between the lines is the right solution. In his video called "learning to drive in 1 day", the start-up reveals behind the scenes of this experience.

Four episodes of exploration follow one another. The car makes mistakes, it is corrected. Then, a first evaluation takes place. The car manages to cover a few meters, then it leaves the road … New episodes of exploration follow one another, with an optimized driving thanks to the returns made by the driver. After 11 test sessions only, the car remains on the road alone.

The "super-driver", the future of the automobile?

to prove this project? What are the possibilities of this approach? The idea is to build an artificial intelligence that, helped by good drivers, could give birth to an autonomous car whose skills would be superior to any driver, the most careful of it. A "superconductor."

DeepMind, the Google subsidiary specializing in artificial intelligence, has shown that deep reinforcement learning methods can lead to exceptional performance. We saw it with the game of Go or chess, the artificial intelligence surpbading the man. Wayve wants to do the same with the car. " We show here that a similar philosophy is applicable in the real world, and in particular with autonomous vehicles. A crucial point to note is that DeepMind's game algorithms required millions of attempts to solve a task. It is remarkable that we have taught a car to follow a direction … in less than 20 tries! "

Modified on 13/07/2018 at 15h17