Helm.ai raises $13M on its unsupervised learning approach to driverless car AI

Four years ago, mathematician Vlad Voroninski saw an opportunity to remove some of the bottlenecks in the development of autonomous vehicle technology thanks to breakthroughs in deep learning.

Now, Helm.ai, the startup he co-founded in 2016 with Tudor Achim, is coming out of stealth with an announcement that it has raised $13 million in a seed round that includes investment from A.Capital Ventures, Amplo, Binnacle Partners, Sound Ventures, Fontinalis Partners and SV Angel. More than a dozen angel investors also participated, including Berggruen Holdings founder Nicolas Berggruen, Quora co-founders Charlie Cheever and Adam D'Angelo, professional NBA player Kevin Durant, Gen. David Petraeus, Matician co-founder and CEO Navneet Dalal, Quiet Capital managing partner Lee Linden and Robinhood co-founder Vladimir Tenev, among others.

Helm.ai will put the $13 million in seed funding toward advanced engineering and R&D and hiring more employees, as well as locking in and fulfilling deals with customers.

Helm.ai is focused solely on the software. It isn't building the compute platform or sensors that are also required in a self-driving vehicle. Instead, it is agnostic to those variables. In the most basic terms, Helm.ai is creating software that tries to understand sensor data as well as a human would, in order to be able to drive, Voroninski said.

That aim doesn't sound different from other companies. It's Helm.ai's approach to software that is noteworthy. Autonomous vehicle developers often rely on a combination of simulation and on-road testing, along with reams of data sets that have been annotated by humans, to train and improve the so-called "brain" of the self-driving vehicle.

Helm.ai says it has developed software that can skip those steps, which expedites the timeline and reduces costs. The startup uses an unsupervised learning approach to develop software that can train neural networks without the need for large-scale fleet data, simulation or annotation.

"There's this very long tail end and an endless sea of corner cases to go through when developing AI software for autonomous vehicles, Voroninski explained. "What really matters is the unit of efficiency of how much does it cost to solve any given corner case, and how quickly can you do it? And so that's the part that we really innovated on."

Voroninski first became interested in autonomous driving at UCLA, where he learned about the technology from his undergrad adviser who had participated in the DARPA Grand Challenge, a driverless car competition in the U.S. funded by the Defense Advanced Research Projects Agency. And while Voroninski turned his attention to applied mathematics for the next decade - earning a PhD in math at UC Berkeley and then joining the faculty in the MIT mathematics department - he knew he'd eventually come back to autonomous vehicles.

By 2016, Voroninski said breakthroughs in deep learning created opportunities to jump in. Voroninski left MIT and Sift Security, a cybersecurity startup later acquired by Netskope, to start Helm.ai with Achim in November 2016.

"We identified some key challenges that we felt like weren't being addressed with the traditional approaches," Voroninski said. "We built some prototypes early on that made us believe that we can actually take this all the way."

Helm.ai is still a small team of about 15 people. Its business aim is to license its software for two use cases - Level 2 (and a newer term called Level 2+) advanced driver assistance systems found in passenger vehicles and Level 4 autonomous vehicle fleets.

Helm.ai does have customers, some of which have gone beyond the pilot phase, Voroninski said, adding that he couldn't name them.