Rather than a 3-D map, an army of sensors and cameras, and whatever else it is that autonomous vehicles use in order to safely navigate the road ahead, researchers at Wayve outfitted their self-driving vehicle with a “deep convolutional neural network” and one monocular camera. The UK-based company then put the car on an empty road and taught it to lane-follow through “reinforcement learning,” which is basically trial and error.
As seen in the video below, the car starts off veering out of its lane almost immediately and is corrected by a safety driver. During each training episode, the algorithm is rewarded for the distance it travels before the driver corrects its course. After 11 episodes, the car travels the entire length of the road without any driver intervention.
This method eliminates the need for “massive models, fancy sensors and endless data,” as the company says, and is the first case of a car driving itself using reinforcement learning. The next step is to add more complex driving scenarios such as roundabouts and intersections.