Nikolai Ensslen

Advancing Physical AI to mastery
Co-Founder & CEO at Vesoma
Co-Founder of Synapticon

[Nikolai Ensslen @nikensslen]

...and former Tesla Optimus Gen3 designer Marcelo Gutierrez, former DeepMind researchers Roland Hafner and Oliver Groth, Etergo (> Ola Electric) founder Bart Rosier, alongside more folks previously with Tesla, Amazon, 1X, Foundation, Neura, Agile, BMW, RobCo, Enchanted and other powerhouses. 😊 Bear with us, this team will make robotic labor a reality.

[RoboIndex @roboindexio]

One of Germany's newest humanoid companies is hiding a surprisingly heavyweight team.

The Munich-based stealth company behind this walking prototype has now revealed Kyle, its first humanoid development platform.

The robot was built from scratch in roughly three months, took its first steps in June, and is now walking using a combination of kinematic planning and self-learned dynamic tracking and stabilization.

Behind it is a very strong team: Synapticon co-founder Nikolai Ensslen, Easybrain co-founder Peter Skoromnyi, former Google DeepMind Controls lead Martin Riedmiller, and former DFKI robot-learning researcher Boris Belousov.

The company has already grown to around 50 people across Munich and Limassol.

Kyle is only the development mule. Generation 1 is scheduled to be unveiled in early 2027.

With Agile Robots already building Agile ONE and Devanthro developing Robody, Munich is becoming an increasingly interesting humanoid robotics cluster in Europe.

🎥 Credits: @nikensslen / @_bbelousov

[Nikolai Ensslen @nikensslen]

@Nicolas_Keller Neat launch, congratulations neighbors! Can’t wait for this to be expanded to full humanoids 🙂

[Nikolai Ensslen @nikensslen]

@JHochderffer @rohanpaul_ai Yes, especially because they will cost 10x as much as they can’t be produced at scale

[Nikolai Ensslen @nikensslen]

@rohanpaul_ai Sure, that’s why we’re redesigning factories and homes for decadesx already to fit the robots that do exist since 🫣🫣🫣 This won’t age well.

[Nikolai Ensslen @nikensslen]

@zobotics Sure, and all these purpose-built robots will each have a scale of…? Like today, because they exist for decades.

[Nikolai Ensslen @nikensslen]

@DJiafei And, I mean, MolmoAct is already hyper slow vs what you desire a robot to be

[Nikolai Ensslen @nikensslen]

@chris_j_paxton Resonance

[Nikolai Ensslen @nikensslen]

@l3d1c Wrong company

[Nikolai Ensslen @nikensslen]

@Scobleizer It’s significantly less than 10y if we do the right things. Currently everybody‘s still so imitation learning pilled, and I start to lose understanding why - the ceiling is very obvious. So is the solution, but it’s harder to build. Onwards.

[Nikolai Ensslen @nikensslen]

@antopatrex1 Absolutely. VLM/As and WAMs are there for language and world understanding and reasoning, but will continue to fundamentally fail in physical mastery. Only self-learning on the real hardware (in conjunction to simulation) can get us to generalizing 99.9% success rates.

[Nikolai Ensslen @nikensslen]

@BrianBellia @RoboStrategy With pleasure. @GoingBallistic5 Scott was one of our earliest visitors beginning of this year, but he didn’t get to see any of our HW because we were still in early stealth and we didn’t want to confuse him with our boxy alpha platform. 😉

[Nikolai Ensslen @nikensslen]

Vesoma. Today the name is public.

Ve for versatile, soma, Greek for body. Then we noticed the same Ve opens vera (truth) and veda (knowledge). Trustworthy, intelligent, versatile embodiments. An AI lab that ships a humanoid.

Its intelligence learns from doing the work, not from watching us do it.

Learn more: https://t.co/3hcIpyyLUp
@vesoma_ai

[Nikolai Ensslen @nikensslen]

Yeah, would‘ve loved to break the black visor scheme, issue is that fixed RGB-cam eyes are creepy too (see NEO) and we actually need more than a dozen cameras in the head, predominantly for safe human detection. Exposing this would render an insect head and a dark visor is the only thing physically available to handle that. That said, our ID team did a spectacular job in shaping the head to be friendly rather than alien. Stay tuned.

[Nikolai Ensslen @nikensslen]

@BrianBellia No worries, I got that 🙂 I would categorize anything that’s „more“ than human superhuman, incl. slip ring joints (btw they are very expensive)

[Nikolai Ensslen @nikensslen]

I have watched a lot of humanoids walk over the last years. In my honest assessment, this is one of the better gaits out there - and it is still our very early stage. Congrats Martin Riedmiller & team!

We decided showing the alpha platform unblurred for the first time. Until now we covered it in every clip, because it is an internal development mule and not the product. Generation 1 is progressing well and we are relaxed about showing what came before it.

And we simply grew fond of this machine. The boxy shape is the direct result of rapid prototyping that allowed for the record design and build time. Exactly that raw frame is what makes the contrast so striking once it starts moving. A boxy machine man with a very organic character.

Generation 1 gets unveiled early 2027. If you’ve build humanoids before or shipped hardware at scale, consider joining.

[Nikolai Ensslen @nikensslen]

@_krishnasuresh Prefer to focus on really learning such skills instead of wasting time finding out which trajectory-object combinations lead to coincident success when playing back recorded trajectories. The amount of unexplainable hope in imitation learning is flabbergasting.

[Nikolai Ensslen @nikensslen]

@MoonWalkerLD @_jan_olsson_ One can build robots as well here 😁

[Nikolai Ensslen @nikensslen]

@cbed32 @_jan_olsson_ Nope. There is nothing in Zurich compared to Munich. Banks, US hyperscaler offices that round-robin the AI talent, and a couple of interesting startups, but not much at scale.

[Nikolai Ensslen @nikensslen]

@_jan_olsson_ 💯 Garching is Munich‘s Palo Alto

[Nikolai Ensslen @nikensslen]

@zhaoran_wang @xfromhome Data you can buy (aka supervised learning) won’t get us there. Machines will need to learn these themselves.

[Nikolai Ensslen @nikensslen]

@zhaoran_wang That might be true, but to be actually useful we need to run them on edge. And physical controls still aren’t and won’tbe solved by these.

[Nikolai Ensslen @nikensslen]

@AIQuanting @humanoidsdaily While aiming to cover all of physical labor long-term, including in homes, we start with high-value use cases across manufacturing and logistics, such as parts commissioning, manufacturing line loading and machine tending.

[Nikolai Ensslen @nikensslen]

@BrianBellia So far we aim for human parity and to make it trustworthy for humans. Once this is achieved, we might look into ways to make it superhuman.