Automakers didn’t build the self-driving car: Google did. That’s a big problem for them.
Hoping to catch up, Ford, Toyota, and Volkswagen are betting on academics. Along with Nvidia, Samsung, Qualcomm and Panasonic, they’re each giving $300,000 to the University of California to fund artificial intelligence research.
The alliance, called DeepDrive, is a rare moment of AI cooperation among car companies, which are racing one another to create the kind of brains that propel Google’s prototype gumdrop-shaped vehicles around Mountain View.
It also highlights the new position universities find themselves in. Their AI lab work is in high demand, and corporations don’t want to wait months or years to get their hands on it.
The companies’ money will go to projects selected by the University of California. In return, the automakers get to give feedback on research proposals; meet the academics toiling away on the tech; and, thanks to the upfront payment, can commercialize any of the research without having to go through the headache of an additional licensing stage.
“They’ve essentially prenegotiated access to software,” said Trevor Darrell, a University of California professor who leads DeepDrive.
In corporate terms, $300,000 might not seem like a lot of money, but together the donations will back 20 to 30 graduate students a year.
It’s a cheap way for the companies to get a bead on a dangerous, unpredictable future.
“If vehicle manufacturers, five years from now, haven’t been to the drawing boards to figure out how to get self-driving tech into their cars, then those companies will be left out,” said Thilo Koslowski, top car analyst at the research firm Gartner. “It’s that dramatic.”
For the University of California, it’s an opportunity to get funding without having to delay publication at the behest of a sponsor. That’s the normal protocol for corporate-backed research, where companies can ask to review results prior to general publication, delaying publication for months — or more.
Openness has become a big deal in AI as the pace of research speeds up. No one wants to be caught reinventing the wheel (or the car). With DeepDrive, the university “can have open research with no publication restrictions, no lockdown of early review for patenting, so the research can move as fast as it possibly can,” Darrell said.
It also gives the university a way to test its theoretical ideas in the real world, said Pieter Abbeel, a professor and one of the principal investigators at DeepDrive.
“We do all this research, but unless you do a startup, where’s it going to go?” Abbeel said. “These companies, they all have experts in the same topics but might not have the time to do research. But they understand everything we’re doing and can translate it very nicely into their own projects and we can see it in action.”
Through the project, the university’s researchers also could get access to driving data from the companies and be able to run their software on the automakers’ vehicle simulators, letting them test new approaches without risking crashing real cars, he said.
The types of problems DeepDrive’s researchers will tackle read like an index page from a science fiction novel: custom semiconductors for vision systems; software to predict how a pedestrian will behave; AI that can drive in unusual terrain; and techniques that let machines learn from human drivers. It will also fund work on Caffe, a programming framework developed by the university that cuts out some of the gruntwork in writing AI code.