20 July 2026

^{The Microagi team (photo: Microagi)} 

 

Ten months after it was founded, Munich robotics start-up microagi has raised $55mn in seed funding to help industrial companies train and deploy robots using data captured from experienced human workers.

The round was led by Hummingbird Ventures, with participation from Northzone, LocalGlobe, Village Global and Swiss venture capital firm redalpine. Microagi says it believes the deal is the largest seed round raised by a German start-up, although the company has not disclosed its valuation.

The size of the investment is striking for such a young company, but so is the problem investors are backing it to solve. Artificial intelligence models have become remarkably good at processing text, images and video because developers can train them on enormous digital datasets. Robots do not enjoy the same advantage.

Learning how to pick up an unfamiliar component, load a machine or sort irregular objects requires data about movement, timing and physical context. Much of that knowledge exists only in the movements of people who have performed the same task thousands of times.

Microagi wants to capture it before it disappears.

 

From Formula One to factory automation

 

The company was founded in 2025 by five engineers and entrepreneurs: chief executive Bercan Kilic, chief technology officer Nico Nussbaum, Yoan Iliev, Anton Poletaev and Artjem Weissbeck.

Their backgrounds stretch from motor racing and academic research to engineering and company building. According to reporting by Business Insider, Kilic worked as an aerodynamics engineer at Red Bull Racing, joining the Formula One team in 2023. Iliev previously worked as an engineer at Mercedes’ Formula One operation, while Poletaev was a researcher at the Alan Turing Institute.

Nussbaum has an engineering background at RWTH Aachen University, while Weissbeck is a serial entrepreneur. It is a founding team shaped less by conventional industrial robotics than by disciplines in which complex physical systems, rapid experimentation and precise data analysis are central.

Kilic has framed automation as a response to a structural problem facing European and US manufacturers: skilled employees are retiring, fewer younger workers are available to replace them and reshoring production is difficult without a substantial increase in productivity.

“If you run factories, the math is already on your desk,” he told Business Insider. “Your most experienced people retire this decade, and their replacements were never born. Reshoring only works if the robots do.”

Data from the International Federation of Robotics gives that argument some context. China installed 295,000 industrial robots in 2024, accounting for 54% of global installations. Europe installed 85,000, including 26,982 in Germany, while the US installed 34,200.

Germany remains Europe’s largest robotics market and the fifth-largest globally. But the gap with China shows the scale of the automation challenge facing European industry, particularly if the continent wants to rebuild production capacity rather than continue relying heavily on imported goods.

Microagi’s data strategy adds another dimension to that challenge. Its technology is intended not only to automate physical work, but also to record aspects of industrial expertise that might otherwise leave the factory with retiring workers.

 

How Atlas teaches robots to work

 

Microagi describes itself as a robotics deployment company rather than a robot manufacturer. Its Atlas platform is designed to connect human demonstrations, AI training and the eventual deployment of a robot on a production line.

In simple terms, Atlas records how people perform a task, trains a robot control policy from those demonstrations and deploys that policy on suitable hardware.

The process begins inside the customer’s operation. Recording equipment captures experienced workers performing a particular task. First-person cameras and, according to secondary reporting, sensor-equipped gloves can gather information about how people move and manipulate objects. The demonstrations give the system examples grounded in the environment where the robot will eventually operate.

These data are then used to train what roboticists call a control policy: the software that converts what a robot perceives into physical actions. Training takes place partly in simulation, allowing the policy to practise and fail without damaging machinery, products or people.

Once trained, the policy can be deployed on the type of robot best suited to the job. Microagi says Atlas is independent of any particular robot manufacturer or hardware supply chain. That could allow customers to apply its software and training process across different machines instead of becoming tied to one robotics platform.

“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,” Nussbaum told Sifted. “So every month we’re there they pull a little further ahead of their competitors.”

This deployment-first approach distinguishes microagi’s near-term proposition from the more distant promise of a universal humanoid capable of walking into any workplace and performing almost any task. The company is starting with specific operations, collecting data where the work takes place and adapting a suitable robot to perform them.

Microagi has not named the industrial customers using Atlas or published performance figures showing how reliably deployed systems complete particular tasks. Those will become important measures as the company moves beyond its initial funding announcement.

 

When domestic work becomes training data

 

Atlas is supported by Shift, the data-collection network operated by microagi. The company says the network has expanded to more than 20,000 participants in 15 countries who are paid to record themselves performing physical activities in factories and homes.

Shift has also offered free home cleaning in New York in exchange for permission to record the work. As Ars Technica reported in May, cleaners arrive wearing head-mounted cameras and capture first-person footage while performing household tasks. The resulting recordings are intended to help train robots to operate in the less predictable conditions of real homes.

The service makes the economics of embodied AI unusually visible. Customers receive free cleaning; microagi receives data that could be considerably more valuable than the service itself.

Shift says faces, identity documents and other personal information are blurred before footage is uploaded to its cloud. Its operators have reportedly been paid around $20 an hour, while the company said that more than 10,000 participants had collectively earned over $5mn during the first quarter of 2026.

Yet the model also raises questions that extend beyond conventional data labelling. A video recorded inside a home may reveal its layout, possessions, security systems and the routines of the people living there. Anonymising faces does not necessarily remove every sensitive detail from such footage.

Microagi has not publicly identified all the organisations buying or accessing the data, or fully detailed how long household recordings may be retained and licensed. There is also a more uncomfortable question for industrial workers: are they preserving their expertise, or generating the training data that could eventually enable companies to automate their roles?

That tension is likely to become more visible as physical work itself becomes a valuable AI input.

 

Building the data layer for embodied AI

 

Embodied AI refers to artificial intelligence that perceives and acts through a physical body. Rather than generating an answer on a screen, an embodied system must understand its surroundings, plan an action and execute it through a robot.

Errors have physical consequences. Objects, lighting and environments also rarely behave as neatly as they do in simulations.

Much of the current research centres on vision-language-action models. These combine visual perception and language-based instructions with robot actions. A model might receive an instruction to place a component in a container, identify both objects through cameras and calculate the sequence of movements required to complete the task.

The difficulty is generalisation. A robot trained to grasp one component in a carefully arranged workstation may fail when the component is rotated, partially hidden or replaced by a slightly different version. Training a separate system for every object and environment is slow and expensive.

Developers are therefore trying to build models that learn from diverse tasks and transfer that knowledge to new machines and settings. California-based Physical Intelligence, for example, is developing general-purpose robot foundation models trained across multiple tasks and robot configurations. Its research indicates that sufficiently large robotic models can also use first-person human video to improve performance where robot-generated data are scarce.

French start-up Genesis AI is pursuing another closely related route. It is developing a model intended to work across different robot types while collecting industrial demonstrations through workers wearing sensor-equipped gloves. The company has also developed a dexterous robotic hand, giving it a broader hardware role than microagi currently claims.

Within Germany, Neura Robotics is taking a more vertically integrated approach. It develops cognitive industrial and humanoid robots alongside the Neuraverse AI ecosystem and physical training facilities where robots practise tasks under controlled variation. Microagi, by contrast, is betting that the intelligence and deployment layer can remain independent of the machine underneath it.

The distinction is also visible elsewhere in physical AI. As MTN previously reported, Gather AI uses drones and computer vision to collect operational data inside warehouses, while Wayve’s research into embodied AI focuses on how autonomous vehicles learn to interpret and navigate the physical world.

These companies are not direct competitors in every market. But together they are exploring different parts of the same emerging stack: hardware, foundation models, real-world data and the deployment systems connecting them.

 

Funding the move from data to deployment

 

Microagi says it will use the seed capital to finance customer deployments with large industrial companies and expand its teams in Munich, Zurich, London and Istanbul.

Zurich has been selected as its global robotics research headquarters, reflecting the city’s proximity to ETH Zurich, specialist robotics talent and one of Europe’s most concentrated industrial regions.

The company is also running a six-month research fellowship offering selected participants access to robotics hardware, human-demonstration datasets, a monthly stipend of $10,000 and up to $2mn in computing resources per fellow. Microagi has not said how much of the seed round will be allocated to the programme or provided a detailed breakdown of its planned spending.

The immediate commercial test will take place inside factories rather than research laboratories. Capturing demonstrations is only the first step. Microagi must show that its policies can convert them into dependable robot performance, remain safe around workers and adapt as production lines change.

Kilic has set an extraordinarily ambitious goal of deploying the company’s technology on 20mn to 30mn robots within five years. That target is better understood as a statement of intent than a forecast supported by current deployment figures.

Yet the underlying bottleneck is real. The next stage of robotics may depend not only on building better machines, but on finding a scalable and responsible way to teach them.

Microagi’s $55mn bet is that the missing instruction manual is already present in factories and homes—in the accumulated movements of the people doing the work.

 

 

Liked this article? You can support our independent journalism via our page on Buy Me a Coffee. It helps keep MoveTheNeedle.news focused on depth, not clicks.

👉 https://buymeacoffee.com/movetheneedle.news