As thousands of dashcam videos on YouTube vividly demonstrate, drivers see the craziest things. Be it an angry bear, a low-flying aircraft or even a guy riding a shopping cart on the freeway, the videos make for entertaining viewing but also illustrate a problem facing developers of self-driving cars: how can you program a computer to make sense of all this?
On Tuesday, chip maker Nvidia introduced a US$10,000 computer that it says will allow cars to learn the right and wrong reactions to different situations, essentially figuring out what to do from experience rather than a rigid set of pre-defined situations.
"Driving is not about detecting, driving is a learned behavior," said Jen Hsun Huang, CEO of Nvidia, during a presentation at the company's GTC 2015 conference in San Jose.
The Drive PX is based on two of the company's Tegra X1 processors and will crunch video from up to 12 cameras. Over time it should learn, for example, to slow down for dogs but not slam on the brakes for a piece of newspaper blowing across the road.
Today's commercial autonomous systems are largely related to detecting when cars stray from their lanes or preventing collisions. Several fully self-driving cars have been developed as part of research projects, but they rely on highly detailed maps and are generally restricted to operating in controlled environments.
A DARPA project already proved the learning technology on a lower level, said Huang. A small autonomous robot was fed with 225,000 images of a backyard. When it started out, the robot ran straight into an obstacle, but after analyzing the images, it managed to successfully scoot around the yard without hitting objects, figuring out for itself how to get around.
The Drive PX is intended to be used by auto makers in research and development projects and is unlikely to mean self-driving cars are coming anytime soon. But if it works as promoted, it could help advance their arrival.
One proponent of autonomous driving, Tesla Motors CEO Elon Musk, said the most difficult part of realizing the technology was at speeds between 10- and 50 miles per hour.
"It's fairly easy to deal with things that are sub five or 10 miles per hour, you just make sure it hits nothing" said Musk, who was speaking alongside Huang at the event. "From 10 to 50 miles per hour in complex suburban environments, that's when you can get a lot of unexpected things happening. Once you're above 50 miles per hour, it gets easier again."
An additional element of Drive PX will ensure that actions learned in one car are shared with others.
Nvidia didn't say which auto makers would be using the platform, which will be available from May, but did say that it's already receiving enquiries from car companies about the technology.
Martyn Williams covers mobile telecoms, Silicon Valley and general technology breaking news for The IDG News Service. Follow Martyn on Twitter at @martyn_williams. Martyn's e-mail address is email@example.com