In the pursuit of creating tomorrow's autonomous machines, a model often requires its own model. Firms that are engineering self-driving vehicles, robots that interact with the physical world, or independent construction machinery amass thousands—even millions—of hours of video for assessment and training. Currently, humans are tasked with organizing and cataloging this footage, requiring them to watch it all. Even at high speed, this approach is not scalable.

Nomadic AI, a startup established by CEO Mustafa Bal and CTO Varun Krishnan, aims to address the challenges faced by clients who have 95% of their fleet data stored unused in archives. The difficulty intensifies when searching for edge cases—the most crucial data often captures rare events that can confuse less-experienced physical AI systems. Nomadic is tackling this issue with a platform that transforms raw video into a structured, searchable dataset using a suite of vision language models. This process enables superior fleet monitoring and the generation of specialized datasets for reinforcement learning, allowing for quicker iteration.

The company revealed an $8.4 million seed funding round on Tuesday, achieving a post-money valuation of $50 million. TQ Ventures led the investment, with contributions from Pear VC and Jeff Dean. This capital will support the onboarding of additional clients and further refinement of the platform. Notably, Nomadic also secured first place in the pitch contest at Nvidia’s GTC conference last month.

“We provide insights into our clients' own footage, from what drives their autonomous vehicles and robots,” Bal explained. “It’s this specific data that propels these autonomous system developers forward, not random information.”

Consider, for instance, the task of fine-tuning an autonomous vehicle to recognize when it may proceed through a red light if directed by a police officer, or identifying every instance a vehicle passes under a particular type of bridge. Nomadic’s platform can pinpoint these occurrences both for compliance needs and to feed them directly into training pipelines. Current clients, including Zoox, Mitsubishi Electric, Natix Network, and Zendar, are already utilizing the platform to advance their intelligent machines.

Antonio Puglielli, VP of Engineering at Zendar, noted that Nomadic’s tool allowed his company to scale its operations much faster than outsourcing would have, and that its specialized knowledge distinguishes it from other market players.

This model-driven, auto-annotation tool is becoming an essential workflow in physical AI. Established data labeling companies such as Scale, Kognic, and Encord are creating AI tools for this purpose, while Nvidia has released a series of open-source models called Alpamayo that can be customized to address the challenge.

Varun Krishnan emphasizes that his company’s offering is more than a simple labeler; it functions as an “agentic reasoning system: you describe what you need and it determines how to locate it.” It employs multiple models to comprehend actions within the footage and contextualize them.

Nomadic’s investors believe the startup’s dedicated focus on this specific infrastructure will lead to success. “The moment an autonomous vehicle company attempts to build Nomadic’s solution internally, they divert attention from their core advantage—the robot itself,” a representative from TQ Ventures commented.

EMBED_PLACEHOLDER_0

The investors also commend Nomadic’s team, highlighting that Krishnan is an international chess master ranked as the world’s 1,549th-best player. Krishnan, in turn, proudly mentions that all dozen or so engineers at the company have published scientific papers.

The team is now deeply engaged in developing specialized tools, such as one that interprets the physics of lane changes from camera feeds, and another that pinpoints the precise locations of a robot’s grippers within a video. The next significant challenge, from the perspectives of both Nomadic and its customers, is to create analogous tools for non-visual data like lidar sensor readings, and to integrate information across multiple sensor types.

“Managing terabytes of video, processing it with hundreds of models each having 100 billion-plus parameters, and then extracting accurate insights is an extraordinarily complex task,” Bal stated.