Deep Learning How to Drive

September 20, 2017

Ryan Gehm

Nvidia's Xavier AI supercomputer chip for self-driving vehicles. (image: Nvidia)

Learning to drive as a teen is a rite of passage, my own experiences indelibly marked in my brain. But the “brain” learning to drive in the future won’t be blowing out sixteen candles, if Nvidia has its way. Deep learning—a programming model that builds a “neural net,” basically a self-adaptive algorithm that acts like a human brain after being trained by data—is the perfect solution for self-driving vehicles, according to Tri Huynh, Nvidia’s Senior Manager of Business Development – Autonomous Vehicles.

Nvidia’s Tri Huynh (second from right) believes that deep learning is the perfect solution for self-driving vehicles. (image: SAE International)

“One of the hardest computer science problems is self-driving,” he said during a Sept. 18 session at SAE COMVEC 17 on Vehicle Architectures for Connectivity and Processing. “The things you have to detect on the road, and make the right decision, is a nearly impossible software problem; you can never write enough software to detect everything you see out there.”

Rather, Nvidia is building a supercomputer inside the vehicle. This processor, called Xavier, was developed at a cost of $1 billion, according to Huynh, and its strict purpose is to drive a vehicle. “We’ve taken all we’ve learned for the past four years on AI [artificial intelligence] and self-driving and put it into one chip,” he said. “To give you some idea of its performance, this processor can do 30 trillion operations per second at 30 watts—that’s about the same performance of 180 MacBook Pros. This is what we think it’s going to take, probably multiples of these type of processors, to get to [SAE] Level 4 or Level 5 [automation].”

The tech company has its own test car called BB8—yes, in honor of the Star Wars droid—that’s been trained to drive by “watching” a human driver. Huynh showed a video clip of BB8 skillfully descending a steep, winding street. “With traditional software, you would basically write algorithms: detect the sign, detect the lanes, don’t hit this, don’t hit that. What we’ve done is mounted a camera inside looking at the person, some sensors looking at the steering angle, and a camera outside looking at the environment. It’s learning just like how you’d teach your kids how to drive; there’s no additional software detecting the curb, the bush, etc.”

Deep learning and AI are being used for non-self-driving situations, too—for example, employing AI as essentially an active safety element in vehicles. In a scenario shown on screen at COMVEC 17, the driver doesn’t see a truck about to run a red light. As she begins to accelerate through the intersection, her car sees what’s about to occur and prevents the accident. “This is a great application for deep learning and AI, for safety in the vehicle,” he said.

The other use case involves California-based Blue River Technology, a precision-agriculture tech company that uses Nvidia technology for its advanced spraying equipment. “They put cameras and our computers on the tractor and they’re using deep learning algorithms to detect what are weeds and [determine] how and when to use pesticide,” Huynh said. The result is a reduction of chemical usage by 90%. Blue River was just acquired in early September by Deere & Company for $305 million.

For self-driving vehicles, AI will provide a base level of performance out of the box, according to Huynh, “but it may need more time to learn your behavior.” Only the good driving behaviors, I hope.