This post is also available in: heעברית (Hebrew)

Network World brings us news of an electric car that teaches itself to see and steer. An engineering team at Nvidia – known primarily for their graphics processing units (GPUs) – has built a self-driving car with only one camera and one Drive-PX embedded computer that uses only 72 hours of training video data to achieve impressive results.

DAVE2, as the project is dubbed, is named after a US Defence Advanced Research Projects Agency (DARPA) programme call DARPA Autonomous Vehicle (DAVE). That programme is already ten years old, but it was one of the first major applications of neural networks to autonomous vehicles. DAVE2 honours that legacy, but exploits the tremendous advances in GPU technology in recent years for a project that is now much more economically feasible.

The Nvidia team trained a convolutional neural network (CNN) to observe a human driver’s actions on the road, and to learn from his driving. In the learning stage, the system used three cameras and two computers to capture and analyse three-dimensional images to learn from, but the actual driving mechanism uses only one camera and one computer.

Once the CNN performed well in simulation, testing was thrown into high gear and unleashed on the roads of London. A human driver supervised on-road testing, in which the computer’s performance significantly improved. Using this method, it took less than 100 hours to train the system to drive autonomously in diverse and complex lighting and weather conditions.

The system is not road-ready yet, as Nvidia writes in the paper, but you can watch the demonstration of the work in progress in the video below: