Nomadic makes video data searchable for training AI models after raising $8.4M in funding

NomadicML Inc. today announced that it has raised $8.4 million in seed funding to address critical video data management issues unique to autonomous robots and vehicles. The round was led by TQ Ventures, with participation from Pear VC, and angel investors from Jeff Dean and several executives from OpenAI Group PBC and Google LLC’s DeepMind.

The startup has developed a platform aimed at allowing robotics and self-driving car developers to search large amounts of video data generated by their vehicles.

These 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 currently have. However, there are some problems that cannot be easily solved. A single car or robot can generate terabytes of data every day, and thousands of fleets are swimming in petabytes of video footage. This unstructured data is generally archived, and most of it is left untouched because it is impossible for humans to physically review and label everything.

Model development is hampered by the inability to take advantage of this rich video data. For example, a model that operates self-driving cars requires more data for important edge cases, such as navigating a construction site during rainy weather, to improve performance. The company likely has hours of video footage of its vehicles doing this in the real world, but the challenge is finding it among those petabytes of archive files. It’s a well-worn adage, but it’s like looking for a needle in a haystack.

Co-founder and CEO Mustafa Bar (pictured, left) said finding this critical data is especially difficult for startups. “Teams are sitting on a treasure trove of video and sensor data, but most of it never becomes usable training signals,” he said.

Nomadic’s solution is a “visual data engine” that turns hours of raw, unorganized video footage into a searchable library of production-ready AI training data. To do this, we rely on advanced computer vision models. These models literally power autonomous teams of agents that monitor video on behalf of businesses and catalog everything for easy discovery. Instead of spending months searching through archives, engineers can query video archives in natural language to find relevant material in seconds.

Chief Technology Officer Varun Krishnan (right) said the system is more than just a data labeling tool. He describes it as an “agent reasoning system” designed to figure out the context and actions taking place within a video, allowing the footage to surface when a human tries to explain it. This allows us to support more complex queries, such as any incidents where a robot gripper interacts with a specific type of object.

The funding will enable Nomadic to expand its platform and meet growing demands from robotics teams drowning in data archiving.

According to TQ Ventures co-founder Andrew Marks, Nomadic is building important capabilities that will help determine who rules the physical AI world. “Physical AI will be won by the team that can learn the fastest from the real world,” he said. “Nomadic gives robotics and AV builders the most practical way to understand their data.”

Photo: Nomadic