Helm.ai reaches $70 million in commercial deals signed for its foundation models

Helm.ai said today that it has signed $70 million in commercial contracts for its physical AI core models over a 12-month period. These contracts included global automotive OEMs, tier-one suppliers and industrial automation companies.

The Redwood City, California-based company sees the automotive industry as its flagship, which involves deep partnerships. But it also has partnerships with clients in the mining and construction sectors.

“These kinds of projects getting to that level of maturity are extremely exciting for us, because they show that there is real demand for the technology that we have built,” said Vladislav Voroninski, founder and CEO of Helm.ai The robot report. “It’s clearly about getting past that threshold to actually being implemented in the real world,”

“Additionally, as a company, we are now on a path to breakeven, which is rare in this industry, and is a testament to the capital efficiency of our approach,” he said.

Founded in 2016, Helm.ai’s software already covers SAE level 2 across 4 autonomous vehicle (AV) programs, sensing production routes in heavy industry, and expanding robotics development.

Inside Helm.ai’s “deep teaching” methodology

Helm.ai trained its core models using its unsupervised “deep teaching” methodology to master the structure of the physical world itself. This separates the problem of understanding an environment from the problem of acting in it.

“The way we approach the problem is different from thinking about it in that there’s some sensor data and then a decision comes out,” Voroninski said. “We actually have structure across the stack, so every part of the stick is still learned end-to-end using DNNs [deep neural networks]. The way we think about solving the problem is, first, to learn to perceive the world, and then to actually act on that perception. So it’s similar to the way humans learn.”

Voroninski likened it to a teenager learning to drive. That teenager doesn’t have to drive for millions of hours to experience every single possible scenario he might encounter on the road.

“They can already sense everything perfectly and make predictions about what other vehicles or people might do, without relying on driving data,” Voroninski said. “This is pretty fundamental and gives us data efficiency, which is really important in the autonomous driving space, but it’s actually even more important in robotics. I would say it’s essential in robotics.”

An automaker simply needs to figure out how it will step up data collection. “For robotics applications, no one has a large fleet of robots yet. There is no data available to even train them this way,” Voroninski said. “So, data efficiency becomes very, very important.”

“We contain our stack in perception, and then everything downstream of perception,” he continued. “This plays an important role in data efficiency and security certification.”

The result is a system that learns from a fraction of the data, generalizes to environments it has never encountered, and implements within the computational constraints of real-world physical systems, according to Helm.ai.

AVs provide Helm.ai with a foundation to adapt to different environments

Voroninski said an important aspect of Helm.ai’s technology is its independence from the environment. The company’s roots lie in autonomous driving, which already requires generalization across a number of different environments. The system may have to handle a busy city street or a deserted street. This provided a good basis for generalizing across multiple environments, he pointed out.

“We have demonstrated that our technology can be generalized to completely different application areas,” Voroninski said. “We can take that same perception stack and use it for autonomous driving purposes, an open-pit mine, or other types of industrial environments.”

So far, Helm.ai has projects targeting production in the AV, mining, and construction sectors. It has also applied its AI technology to delivery drones.

“When we train baseline models, we don’t just limit ourselves to driving data or a specific target area,” Voroninski said. “Our baseline models are trained pretty generally on datasets well beyond those specific applications, and it’s similar to how a human would learn how to actually do these things.”

Helm.ai has learned other things from the automotive industry. Autonomous driving comes with real-time latency constraints and stringent safety regulations. Being able to handle this area makes the company well equipped for others, Voroninski said.

Helm.ai’s software is robot-agnostic

Voroninski discussed using Helm.ai technology on autonomous vehicles, drones, humanoid robots and more. While it may seem difficult to generalize to many robot form factors, Voroninski said that these different form factors all present similar problems.

“The problem of perception starts with the sensors you’re using and how they’re configured,” Voroninski said. For AVs, this typically means a 360º view with a stack of cameras, lidar and radar.

“We have shown that we can work quite well with these types of sensor modalities and simulate them all at the same time,” Voroninski said. “When you move on to other areas, you’re still talking about sensors sitting on some form factor in some configuration. From that perspective, it’s exactly the same thing.”

“It’s essentially the same problem, just different definitions for what you want to detect, what you want to track, and the kinds of behaviors you’ll be interested in,” he continued.

A person on a construction site might behave differently than a pedestrian on the street. But the same technology can make predictions about both of these people.

Looking ahead, Voroninski said Helm.ai is interested in working with robotics companies across all industries and embodiments.

“Where robotics is today is almost where autonomous driving was 10 years ago,” Voroninski said. “It’s really entering its inflection point, so we’re really excited to take the things we’ve learned from the commercial traction we’ve gotten and apply them to a lot of different areas.”