Nomadic analyzes video data to train AI models after raising $8.4M in funding

NomadicML Inc. said today it raised $8.4 million in seed funding to address the critical headache of discrete video data management for autonomous robots and vehicles.

The round was led by TQ Ventures and saw participation from Pear VC, with Google LLC intelligence wizard Jeff Dean and a number of executives from OpenAI Group PBC and Google’s DeepMind investing as angels.

The startup has created a platform that aims to help automotive and private developers search through the large volume of video data generated by their fleets.

Such companies are racing to develop “physical” artificial intelligence models that will allow robots and cars to operate in real-world environments with far greater autonomy than they do now. But there is a problem that cannot be easily solved. A single car or robot can generate terabytes of data each day, and for those active forces worth in the thousands, that means they’re swimming in petabytes of video footage. This unstructured data is often archived, and most of it sits there untouched, because it is not possible for humans to physically review and label everything.

Not being able to take advantage of this wealth of video data slows down model development. For example, a model using an autonomous vehicle needs more data for critical cases, such as navigating a construction site in rainy weather, to improve its performance.

The company will likely have hours of video of its cars doing this in the real world, but the challenge is finding it among those petabytes of archived files. It’s an old, worn-out saying, but it’s really like looking for a needle in a haystack.

Founder and CEO Mustafa Bal (pictured, left) said getting this sensitive data is very difficult for startups. “Teams are sitting on a gold mine of video and sensor data, but most of it never becomes a usable training signal,” he said.

Nomadic’s solution is a “virtual data engine” that transforms hours of raw, unstructured video into a searchable library of production-ready AI training data. It relies on advanced computer vision models to do this.

Those models empower independent teams of agents who literally watch video on behalf of companies, documenting everything for easy access. Developers can then query their video archives in natural language to find what’s important in seconds, instead of spending months sifting through their archives.

Chief Technology Officer Varun Krishnan (right) said the program is more than just a data labeling tool. He describes it as an “agent-based thinking system” designed to detect the context and actions that occur in a video, in order to display that video when someone tries to interpret it. This is what enables it to support more complex queries, such as every single instance where the robot gripper encounters a certain type of object.

The funding will enable Nomadic to scale its platform and meet growing demand from robotics teams that have found themselves mired in data warehouses.

According to TQ Ventures founder Andrew Marks, Nomadic is building a key capability that will help determine who rules the virtual AI world. “Physical AI will be defeated by teams that can learn quickly in the real world,” he said. “Nomadic gives robotics and AV developers a more efficient way to understand their data.”

Image: Nomadic