Nvidia’s ‘Kitchen Sink’ for AV Testing

May 1, 2018

Sebastian Blanco

Nvidia CEO Jensen Huang laid out the problem during his keynote address at the company’s 2018 GTC conference in San Jose in late March. Autonomous driving technology is “probably the hardest computing problem that the world has ever encountered,” he said.

Huang then announced a new cloud-based system for autonomous-vehicle testing that uses two servers to create a complex virtual environment and then virtually “drives” autonomous vehicles through it. Called Drive Constellation, the system is meant to bridge the gap between the billions of miles of test data needed to get autonomous vehicles to SAE Level 5 capabilities and the limited amount of time these cars can spend testing in the real world.

Drive Constellation blends a simulation of a self-driving vehicle’s sensors (cameras, LIDAR, and radar) that comes from Nvidia’s Drive Sim software in one server, with a self-driving car fitted with the company’s Drive Pegasus high-performance AI computer to process the data. The result is a virtual autonomous vehicle that can cover the ground real cars can’t.

“We call this Drive Constellation,” Huang said. “Pegasus in the Sky. A constellation of them. With just 10,000 constellations, we can cover three billion miles a year.”

That sort of testing is what will give OEMs and regulators the confidence in self-driving vehicles, said Tim Wong, an Nvidia technical marketing specialist.

“All the regulatory guys — and even the car companies — are scratching their heads and asking, ‘How do I know I’ve done enough testing for Level 5 vehicle to put it on the road and let it go?’ But if it’s passed a hundred million tests and we’ve thrown everything and the kitchen sink at it, I’d feel a little bit better,” he said.

Wong’s virtual ‘kitchen sink’ allows for reproducible tests, which is a challenge for on-road testing.

“Engineers hate when they have to get the algorithm right, but every time they test it, it’s going to be different,” Wong asserted. “It’s never going to converge. Having ‘sim’ gives you a way to do reproducible test conditions and then test all these different scenarios. The regulatory people are enamored with this.”

Drive Constellation also will allow Nvidia to do ‘hyperspeed’ testing, which means more miles tested in less time, the company claims.

“That’s 60,000 miles in one hour, or every road in the U.S. in three days,” Wong noted. “We can do that in rain, do that in the summertime, do that in the fog. Then I start to have a better feeling about how good this autonomous car is going to drive.”

It’s not just the engineers who feel better. “The whole hyperspeed thing is a big deal because regulatory agencies, who are very conservative, want to certify a billion miles tested,” Wong said. “Well, good luck with that physically. We’ll see in you 10 years. Drive Constellation gives them a tool to actually realize something safe.”

CEO Huang said that Nvidia’s on-vehicle autonomous drive devices will be both ISO 26262 and ASIL-D certified. The company’s automotive partners will have access to Pegasus in mid-2018. Early-access partners will be able to get Drive Constellation in 3Q18.

---