Nvidia’s Drive AutoPilot System Is for Human-Supervised Self-Driving

Nvidia’s vehicle automation business has expanded its focus beyond fully self-driving Level 5 AVs to a platform designed to improve on current driver-assistance systems next year.

Nvidia Drive AutoPilot, announced at International CES in Las Vegas on January 7, is designed to be more capable than current high-end assistance systems such as Tesla Autopilot but is still intended for cars where a human driver is ready to take over. Nvidia calls it “Level 2+.”

Less than a year ago, at its annual GTC conference, Nvidia heavily promoted its hardware and software for full self-driving. It introduced the Drive Constellation platform for autonomous driving simulation at the mid-March event, which took place just after the fatal March 18 Uber crash in Arizona that shook up the AV industry.

Nvidia remains committed to fully self-driving cars, said Danny Shapiro, Nvidia’s director of automotive, in a pre-briefing on the CES news.

Drive AutoPilot will be able to perform “supervised self-driving” on highways between on-ramps and off-ramps, according to Nvidia. This will include navigating lanes that merge and split, and changing lanes, in addition to common driver-assistance functions such as adaptive cruise control, automatic emergency braking and lane-keeping.

Cars with Drive AutoPilot will also be able to determine their location and make path-planning decisions, the company announced. They can use HD maps where available and learn routes that a driver takes outside of areas with prepared maps.

The system will outperform current ADAS (advanced driver assistance system) platforms, including Tesla’s, because it will use more advanced AI and more powerful hardware, Shapiro said.

“This is a whole different class of computing, a whole different class of software, a different class of experience,” he added.

Vendors can develop their own systems on top of Nvidia’s software stack. At CES, Tier 1 suppliers Continental and ZF are announcing “Level 2+ ” systems based on Nvidia Drive AutoPilot, with plans to enter production next year. Through over-the-air software updates to Drive AutoPilot software, manufacturers will be able to add new features to cars already on the road.

By applying some of the technologies the company has developed in pursuit of fully self-driving cars, such as its Xavier in-car computing hardware and multiple deep neural networks in software, Nvidia has developed a system that won’t fail as often as current ADAS platforms do, Shapiro said. He cited findings last year by the Insurance Institute for Highway Safety, which tested ADAS features in several high-end cars and said some of them frequently failed at key tasks such as detecting other cars and following lanes on curvy and hilly roads. When they did, drivers needed to quickly retake control.

Nvidia Drive AutoPilot will run on the Xavier system-on-chip, with 30 trillion operations per second of processing power, double the capacity of other “autopilot” systems, Shapiro said. It will use the company’s broad set of neural networks, including ones for mapping, identifying other road users, detecting open spaces in traffic and other tasks.

The more powerful hardware and software will allow manufacturers to incorporate five or more cameras, beating current ADAS systems that typically only use three, Shapiro said.

Systems based on the platform can incorporate both Nvidia’s Drive AV and Drive IX software stacks. While Drive AV performs automated driving tasks, Drive IX handles the driver and passenger experience. It can be used for driver monitoring systems that detect drowsiness or distraction, as well as for new types of displays that give drivers access to the car’s surround-view cameras and let them monitor the Drive AutoPilot. AI-powered natural language processing will support voice interfaces in local languages, Shapiro said.

Stephen Lawson is a freelance writer based in San Francisco. Follow him on Twitter @sdlawsonmedia.