Vesoma's first humanoid has a matte-black face and a faint purple glow. It is a render. The company says so in small print beneath the image, on the same page where it promises the machine is real enough to build.
On 22 September the Munich startup left stealth with more than 60 employees, 3,500 square metres of laboratory space across Munich and Limassol, and a first humanoid called Vesoma 1. None have shipped. The first units are being built now.
The pitch is a clean fork in the humanoid race. Stop teaching robots by copying people. Let them learn physical work from their own bodies and their own failures.
It is an early-stage European challenger with unusual pedigree and no disclosed funding round. Today the asset is its people, not a cap table.
For investors the question is timing: does self-learning reach full-shift reliability before imitation-trained fleets scale to fill the same warehouses?
The bet against imitation
Most humanoids today learn the way an apprentice copies a master. An operator performs a task in a motion-capture rig or a teleoperation suit, and the robot imitates the recorded trajectory. The method has produced videos that look a great deal like competence.
Vesoma's founders say those videos mislead. A policy trained on demonstrations learns what the work looks like, not what the work is. It holds while conditions stay inside the recording and freezes when they leave it. Whole-body work on a warehouse floor nobody tidied is precisely the case that leaves it.
Their alternative is to let the robot explore. A self-improving agent tries, fails, and discovers what works for its own frame. Competence that comes from contact with physics should survive a chipped floor and a shifting payload. Competence copied from video, the argument goes, should not.
That is a harder path, and Vesoma is honest about why. Contact-rich manipulation, whole-body balance on hostile ground, tasks where one wrong decision breaks something expensive — all of it has to be learned rather than rehearsed. The company says three conditions must hold at once, and that they rarely do.
The broader market gives the bet its stakes. Warehouses, factory lines and care work are short of hands across Europe, and the shortfall deepens each year as the workforce ages. Physical work is a multi-trillion-dollar input that money is finding harder to buy. Whoever makes it available again sells infrastructure, not a gadget.
Venture funding into humanoid robotics
2026 investment is already roughly double the whole of 2025, which means capital is arriving faster than reliability. · Dealroom, 2026
Pedigree is the part you can verify
The team is unusually strong for a company this young. Nikolai Ensslen co-founded Synapticon in 2011 and spent 15 years building motion control and functional-safety systems for machines; he stepped away from it in December 2025 to start Vesoma. Co-founder Peter Skoromnyi built Easybrain, a logic-puzzle publisher whose apps have been downloaded billions of times. Both men have run companies that reached global scale.
The hire that matters most for the thesis is Martin Riedmiller, Vesoma's chief AI officer. He was a research director at Google DeepMind, and his reinforcement-learning work helped make sample-efficient learning practical on physical systems. Vesoma notes that other labs train with methods its team published, and that other robots run on actuators and safety architecture its team designed.
What the company has not disclosed is money. It began work in December 2025 and has announced no funding round. A team of 60-plus across two countries does not run on goodwill. The silence points either to a raise that has not been papered for public consumption or to backers who prefer to stay quiet. Either way, the absence is a fact to price, not a footnote.
Three conditions Vesoma says rarely coexist
The body has to be engineered to be trained, and to survive being trained.
The team needs enough hardware to iterate weekly rather than quarterly.
Why it builds its own body
Vesoma designs its own kinematics, actuation and sensing instead of assembling industrial parts. Ensslen's argument is that the body is a learning parameter. Actuators chosen for closed-loop training let a policy optimise end to end. Off-the-shelf components optimise for the catalogue, not for the agent.
Vesoma 1 is a 170-centimetre, 75-kilogram machine with 26 degrees of freedom, 12 of them in the hands. Its listed top speed is 4 kilometres per hour and its rated strength is 5 kilograms. The imagery published so far is design visualisation, and the company is upfront about that. The robot in the picture is a rendering. The machines on the floor are prototypes.
The competitive field is crowded and well funded. Figure, Agility Robotics, 1X and a long list of Chinese entrants have raised far more, and several already run pilots with industrial customers. Vesoma is not trying to outspend them. It is trying to argue that the imitation-heavy approach most of them use has a ceiling, and that the ceiling shows up the moment a robot leaves the demo cell.
| Dimension | Imitation learning | Self-learning (Vesoma) |
|---|---|---|
| Training signal | Recorded human demonstrations | Own interaction with physics |
| What it learns | What the task looks like | What the body can do |
| Behaviour off-script | Degrades outside the recorded distribution | Aims to hold as conditions change |
| Hardware coupling | Adapted from industrial components | Designed for the learning loop |
Compiled from Vesoma's published approach and Humanoid.guide specifications, 2026
The sector already has a proof metric, and it is not the one in the launch video. As we wrote in September, the number that mattered in 2026 was deployments that hold across a shift, not demonstrations that hold across a take. Vesoma's bet is a bet about which learning method gets there first.
What does full-shift reliability look like by 2028?
Probability: 40% — the method is credible and the team is strong. No public evidence yet shows self-learning scaling from locomotion to contact-rich manipulation, and no disclosed capital buys the iteration cycle it needs.
✅ Arguments for
Owning the body removes the integration tax that slows imitation-trained fleets.
Confirmation criteria: a named customer publishes a task that a Vesoma unit performs, unattended, for a full shift.
❌ Arguments against
Self-learning has not yet been shown to scale from walking to messy contact work.
Disconfirmation criteria: eighteen months pass with no customer deployment and no disclosed round.
A disclosed funding round, and whether the lead is a specialist robotics or deep-tech fund.
The first named customer, and whether the task is a pilot or a paid deployment.
Published reliability numbers over a shift, not a single take.
Whether rivals copy the self-learning approach or keep scaling imitation.
Development scenarios
🟢 Optimistic scenario (30%)
Implications: the learning method becomes the differentiator, not the hardware, and Vesoma's valuation re-rates.
🟡 Base-case scenario (50%)
Implications: the thesis survives on optionality; the payoff is further out than the headlines suggest.
🔴 Pessimistic scenario (20%)
Implications: the approach was right and early; the equity outcome depends on who buys the team.