Computer chip maker Nvidia made a surprise announcement last night at CES 2015 that it is developing a computer vision-based system for driverless cars, called Drive PX. The company, which is known primarily for making graphics-displaying chips for video games, also showed footage of the system during road tests. Drive PX wasn’t controlling the vehicles, but it did manage to pick out and identify objects in real-time. It differentiated between SUVs, pickup trucks and vans, and identified traffic lights, pedestrians and cyclists. It also recognized that the vehicle behind it with the flashing lights was a police car. It did all of this in real-time, processing footage from multiple cameras angles at once.
If you care about robot cars, this is the time to get excited. What Nvidia is proposing has the potential to accelerate the arrival of autonomous cars by years. Without getting overly technical — as Nvidia arguably did, turning what should have been a super-charged press conference into more of a collegiate lecture — the Drive PX system is a cheaper, faster alternative to the status quo’s (read, Google’s) driverless navigation. While most autonomous driving research relies on interpreting data from a variety of sensors, such as radar, ultrasound and video cameras, Nvidia is focusing solely on video.
Drive PX can theoretically analyze footage from 12 cameras at once, and classify up to 150 objects at any given moment. That ability to draw meaning from the visual world is based on training sessions, where researchers feed millions of images into the system, and either confirm or correct its conclusions. This approach is part of a larger field within artificial intelligence research, called deep learning. Facebook uses deep learning for facial recognition, and Google appears to be using it for just about everything search-related. Nvidia doesn’t have the resources of those or many other Silicon Valley companies, but as of right now, the chipmaker is on track to becoming a major player within the most well-funded branch of AI research. The company is already a long-term partner with BMW, which says its committed to using Nvidia’s technology for its own autonomous research. And if Nvidia is right, and carmakers don’t need wildly expensive laser rangefinders and radar systems (one of the reasons that outfitting a current Google car for autonomous operation costs at least $100,000), then mainstream versions won’t be quite as luxury-priced, for quite as long.
The bad news is that the development of autonomous cars still has to move at the automotive industry’s pace, which is glacially slow compared to software development. Robotic vehicles don’t yet exist as products, and therefore haven’t begun to face the necessarily stringent safety requirements imposed on cars and trucks. For all the hundreds of thousands of test miles that Google has under its belt, no one has sent a driverless vehicle onto black ice, or determined how it actually performs in the moments before and after a serious collision. And some of the safety issues will be entirely new to the industry, and to regulators. During last night’s conference, Nvidia CEO Jen-Hsun Huang talked about one of the advantages of using deep learning for autonomous control. As vehicles throughout the world occasionally came across an object that they couldn’t classify, they could send those error reports to researchers, who would then push out a system-wide update. “It would start off being pretty darn smart, but over time it could continue to learn, just like us,” said Huang. It’s a great idea, but you don’t have to be a cybersecurity expert to recognize the danger of having an entire fleet of robotic cars connected to the internet, waiting for their algorithms to be fundamentally altered.
Luckily, the same factors that can make the automotive appear to move so slowly will most likely prevent anything resembling a Hollywood-style robocar hack, or similar potentially lethal outcomes. The enthusiasm and innovation of companies like Nvidia will be mercifully tempered. So proponents of autonomous driving, and all of its attendant benefits to human existence, should rejoice in the fact that a chipmakers is trying to hitch its wagon to the robotic future of the automotive industry. But there’s no need to remember the name Drive PX, or look for products that run on it. The decision of which self-driving car you should pick is one for another decade.