Microagi nabs $55M to teach factory robots how to work

Munich-based startup Microagi GmbH today announced it raised $55 million in seed funding led by Hummingbird to change how industrial robotics works with artificial intelligence models.

Northzone, LocalGlobe, Village Global and Redalpine also participated in the round. The capital raise arrives around ten months after the company was founded by Formula One engineers from Red Bull Racing and Mercedes-AMG Petronas, two high-octane racing teams.

Microagi is helmed by Bercan Kilic, whose career path includes being a former aerodynamics engineer for Red Bull Racing in 2023. He left his position in the sport to found the AI company.

The power behind Microagi is Atlas, a robotics data and deployment platform for industrial companies, which ingests large volumes of data for training – something that is at a premium for AI-driven robotics – and integrates it to fine-tune frontier robotics models for customers.

The company does not build robots or robotics AI models; instead, it provides a powerful data layer that enables it to teach existing models to do a better job.

“Our partners build genuinely good robots and models,” said Chief Technology Officer Nico Nussbaum. “Our job starts after that, on the factory floor. We put our engineers on site with each customer, and the system learns from their real operations and feeds that back into the next run, so every month we’re there, they pull a little further ahead of their competitors.”

By using the data captured from industry customers, using dedicated recording hardware, and a secure ingestion platform designed to curate training data, the company is able to multiply the data and fine-tune AI models for plant-specific tasks. After deploying to the end customer, the company continues to refine its models using factory-floor data in a reinforcement loop designed to improve accuracy.

The company boasts that its design foundation helps close the gap between impressive robotics demos and actual “hands-on” capabilities in industrial environments.

Although no specific customers have been named, the firm told Business Insider that five companies are currently collecting data through the Atlas program, with one customer preparing to deploy robots in a factory. Customers span industries including automotive, logistics and food.

Robotic data is sparse, so Microagi brought its own

Microagi went viral last month when Shift, the company’s consumer-facing arm that collects data, said it would begin offering New York City residents free home cleaning.

As part of the offer, housecleaning staff will arrive with cameras attached to their bodies. The company’s app website said it would connect “New Yorkers with free, trusted professional house cleaners” so that Microagi could record “first-person-cleaning footage to help train the next generation of household robots.” It adds that there is “no catch” and that the free cleanings will go on for a limited time.

Why does this matter? Although large language models and image models, the best-known AI models on the market, have a giant volume of text and images produced directly by human hands to use for training – robot AI models lack a lot of data they can use for training. This is because although there’s a lot of video displaying people on YouTube and other sources, there is a paucity of video of people actually doing things, such as folding clothing, sorting objects, cleaning and other activities.

Not just any video will work either. Most robotics platforms must be trained on video that has been curated for best practices, good lighting, full view of arms and hands, and other conditions that make it easier to turn into movement telemetry. This means that many robotics AI model developers must gather their own video and movement data from other sources.

In this case, Microagi is offering spotless apartments for the side benefit of useful domestic data.

The company also announced that it’s launching private chefs in San Francisco with the same objective: to provide data for robotic cooking. The video announcing the project is an exercise in secondhand awkwardness, but it gets the point across.

Image: SiliconANGLE/Microsoft Designer