Introducing Vesoma

Where we are

Vesoma was registered and began work in December 2025. Since then we have grown to more than 60 people, fitted out over 3,500 square metres of lab and office space in Munich and Limassol, and designed, built and walked our first humanoid prototypes. From nothing to a walking machine of our own design, moving on a policy it learned itself, in six months. Today we are heads down working on software, AI and Vesoma 1.

Speed matters here for a specific reason. Reliability in physical AI is not designed in one pass. It is reached by iterating on real hardware, again and again, and every week saved inside that loop compounds into the next one.

The problem

The work we all depend on is losing the people who do it. Night shifts in warehouses. Monotonous and hazardous work on factory lines. The lifting and turning that fills care work, leaving the people who do it too little time for the part only a person can give.

Most people already feel the shortage without naming it. The parcel that takes a week. The repair booked six weeks out. House chores that we need to do ourselves. Each of these is the same shortage surfacing in ordinary life, and it will deepen: the people who would take these jobs over the next two decades have already been born, and there are not enough of them.

Conventional automation has taken the work that repeats exactly. What remains requires the autonomy, mobility, dexterity and adaptivity of a person.

It is hard to name a larger unsolved problem, or a larger opportunity. Physical work is the foundation of every economy, worth trillions annually, and it is the input money is finding harder and harder to buy. Whoever makes it available again will have built something closer to infrastructure than to a product.

Why it has not been solved

There is no shortage of humanoids. There is a shortage of humanoids that work.

The demonstrations are frequently impressive. We have yet to see one meet the standard a customer actually sets: the same task, performed correctly, across a full shift, in an environment that was never designed to be orderly, without an engineer on hand to intervene.

The prevailing bet is that enough human demonstrations will close that gap. A machine trained this way learns what the work looks like rather than what the work is. It performs well until conditions leave the distribution it was shown, and then it fails without knowing that it has, which is a failure mode no workplace can accommodate. The situations that matter most are precisely the ones nobody thought to record.

Our approach

We are building a self-improving physical agent that learns from its own interaction with the world. It explores, it fails, it discovers what works for its body, and it improves the longer it runs. Competence acquired this way is grounded in physics rather than in examples, which is what makes it hold when conditions change.

The walking in our lab is an early instance. Nobody programmed that gait. The system was given a body and an objective and worked out how to move for itself, in ways no engineer would have specified.

Locomotion is where this method has proved itself, and it is also where the industry has left it. Almost everyone building humanoids is training the rest of the machine on human demonstrations instead, because carrying self-learning further is genuinely hard: contact-rich manipulation, whole-body work in environments nobody tidied, tasks where a wrong decision breaks something. Yet, we are convinced it is the route to reliability, and we intend to take it the whole way.

Three things have to be true at once. The learning has to be efficient enough that a physical machine reaches competence in a realistic number of attempts. The body has to be engineered to be trained, and to survive being trained. And one team has to hold both, because every change to the body changes what the agent can learn, and most of the limits an agent runs into are hardware decisions made months earlier.

Those conditions rarely coexist. Algorithms written by the people now building this company helped make reinforcement learning data-efficient enough to run on physical systems, and parts of the field train with their methods today. Scaling that from locomotion to physical mastery is the work in front of us.

Why we build the body

The agent is designed to reach mastery across tasks and environments. Generalisation is the entire point. We build our own humanoid because owning the body is what makes the learning fast: kinematics we can optimise a policy against end to end, actuation and sensing designed for closed-loop training rather than adapted from industrial components, and a machine capable enough that the ceiling is the agent and not the hardware.

The human form is deliberate. It is the shape every workplace was built around. It is the form factor with the clearest path to automotive-scale volume. And it is the form people read most easily, which matters more than it sounds when the question is whether someone will work alongside it.

Safety

Safety guides our design from the first sketch rather than the last review. It determines how the system moves, how it decides, and what it does when conditions fall outside what it was built for.

This is the part of the work that cannot be shown in a video, and it is the part that decides whether a humanoid is deployable at all. Our co-founder developed the first safety architecture for humanoids, which the industry is now adopting. We build to the standards this category will be held to, which in several cases means helping to write them.

Trust

A humanoid is not equipment that is installed and forgotten. It operates alongside people, and the trust that requires is earned by being legible rather than familiar: clear about what the system does, how it behaves, and where it stops.

Data follows from the same principle, so we state the arrangement plainly. Our machines' experience is the raw material our models learn from, and what they learn in one deployment goes on to improve every machine, everywhere. What moves between sites is capability, not a customer's data: operational data remains confidential between the customer and us. Every deployment benefits in turn from all the others.

So does longevity. Systems that people depend on are built to last, to be serviced, and to justify their cost over years rather than quarters. These are the standards Europe holds technology to, and we expect the rest of the world to arrive at them for this category too.

Civil only

We are building a helping hand, an AI in the body essentially. With that comes a responsibility to ensure that it does not cause harm. So the commitment was made once, at the founding: not for weapons, military, police, or surveillance. Directly or indirectly.

Why we build it in Europe

Europe industrialised productivity technology before anyone and has stayed competitive at the top of it since. The machine tools, drives and automation lines that equip factories worldwide come disproportionately from Europe, and no region has lived longer with automation at scale, or learned more about what it takes for a machine to earn its keep for decades.

A humanoid is judged on what it can do, how well it does it, and how long it keeps doing it. Those requirements pull against each other, and holding all three at volume is the engineering problem European industry has spent a century solving, in machine tools, in automotive, in the automation that runs both. It is the same problem, in a new form factor.

Members of our team have built hardware inside the fastest-moving manufacturing ecosystems in the world and know precisely how that speed is produced. We are matching it while leaving behind the habits that earned European engineering its reputation for being slow and complicated. Going from nothing to a walking machine in six months is the evidence we can offer so far.

Our team

Other laboratories train with reinforcement learning methods our team published. Other robots run on actuators, motion controllers and safety architecture our team developed. We have built consumer products downloaded billions of times and companies that grew to global scale. The combination is unusual, and it is the reason to expect this to work.

Munich and Limassol

Munich is one of the few places in the world where AI researchers and industrial hardware engineers can be hired in the same building, and where engineers who have taken complex machines from prototype to production are the norm rather than the exception. Limassol is one of the fastest-growing hubs of the European digital economy.

Both are places we build in. That is deliberate. A company intending to supply labour infrastructure to Europe should be European from the ground up, rather than the European subsidiary of something headquartered and controlled elsewhere.

The plan

1. Build an agent that learns physical work from its own experience rather than from imitation.
2. Build the body that makes that learning fast, and own it end to end.
3. Reach reliability on real work, in real conditions, at the standard our customers set.
4. Deploy where the need is most acute, and let every deployment train the next one.
5. Widen the range of work one platform can do until the availability of hands no longer limits what gets built, made and cared for.

The name

Ve for versatile. Soma, the Greek word for body. A body that can do many things - and in what we build, the body is not the packaging around the intelligence. It is where the intelligence comes from.

Boilerplate

Vesoma is a European physical AI company founded in December 2025 by Nikolai Ensslen, co-founder of the motion-control and functional-safety company Synapticon, and Peter Skoromnyi, co-founder of the logic-puzzle games company Easybrain. Vesoma builds physical general intelligence: humanoid systems whose agent and body are developed as one, learning new tasks through real-world experience. AI research is led by Chief AI Officer Martin Riedmiller, formerly Research Director at Google DeepMind. Vesoma operates from Munich, Germany, and Limassol, Cyprus.