To build future autonomous machines, you may need a model.
Companies developing self-driving cars, robots that manipulate physical environments, and autonomous construction equipment are collecting thousands, if not millions, of hours of video data for evaluation and training purposes.
It’s now a human job to organize and catalog that video, and humans have to watch all the videos. Even if I fast forward, the scale doesn’t match. NomadicML, a startup founded by CEO Mustafa Bal and CTO Varun Krishnan, wants to solve a problem for customers where 95% of their fleet data is stored in archives.
This challenge becomes even more difficult when looking for edge cases. The most valuable data represents events that occur rarely and can confuse inexperienced physics AI models.
Nomadic is working to solve this problem with a platform that transforms footage into structured, searchable datasets through a collection of visual language models. This improves fleet monitoring and enables creation of unique datasets for reinforcement learning and faster iterations.
The company announced an $8.4 million seed round on Tuesday at a post-money valuation of $50 million. The round, led by TQ Ventures with participation from Pear VC and Jeff Dean, will allow the company to acquire more customers and continue to improve its platform. Nomadic also won last month’s Nvidia GTC pitch contest.
The two founders, who met as undergraduate computer science students at Harvard University, “kept running into the same technical challenges over and over again in their work” at companies like Lyft and Snowflake, Bal told TechCrunch.
tech crunch event
San Francisco, California
|
October 13-15, 2026
“We’re giving people insight into their AVs (and) whatever drives their robots, their footage,” he said. “That’s what moves autonomous system builders forward, not random data.”
For example, imagine fine-tuning an AV’s knowledge that it can run a red light in response to a police officer’s instructions, or isolating a vehicle whenever it passes under a certain type of bridge. Nomadic’s platform allows you to identify these incidents for compliance purposes and feed them directly into your training pipeline.
Customers such as Zoox, Mitsubishi Electric, Natix Network, and Zendar are already using the platform to develop intelligent machines. Antonio Puglielli, vice president of engineering at Zendar, said Nomadic’s tools allow the company to scale up work much more quickly than outsourcing, and that its domain expertise differentiates it from other competitors.
This type of automated model-based annotation tool is emerging as a key workflow for physics AI. Existing data labeling companies such as Scale, Kognic, and Encord are developing AI tools to do this work. Meanwhile, Nvidia has released Alpamayo, a family of open source models that can be adapted to address this problem.
Varun insists that his company’s tools are more than labelers. It’s an “agent inference system. You describe what you want, and it figures out how to find it.” It uses multiple models to understand the action that’s happening and place it in context. Nomadic’s backers hope startups that focus on this particular infrastructure will emerge victorious.
“It’s the same reason Salesforce doesn’t build its own cloud and why Netflix doesn’t build its own[content distribution facility],” Schuster Tanger, a partner at TQ Ventures, who led the round, told TechCrunch. “The moment self-driving car companies try to build Nomadic in-house, they get distracted by what will make them win: the robots themselves.”
Tanger praised Nomadic’s talent, noting that Krishnan is an international chess master ranked 1,549th in the world. Meanwhile, Mr. Krishnan boasts that all of the company’s dozen or so engineers have published scientific papers.
They are currently hard at work developing specific tools, such as tools to understand the physics of lane changes from camera footage and tools to derive more precise locations of robot grippers in videos. The next challenge from Nomadic and its customers’ perspective is to develop similar tools for non-visual data, such as LIDAR sensor readings, or to integrate sensor data across multiple modes.
“Juggling terabytes of video and matching it against hundreds of 100 billion-plus parameter models to extract accurate insights is extremely difficult,” Bal says.
