The 4 Layers Of Brain Behind One Of Europe’s Leading Humanoid Robots

Humanoid AI, a UK-based firm, is emerging as a significant European player in the humanoid robot market. Their HMND 01 robot primarily features a wheeled design, prioritizing industrial applications due to greater energy efficiency, stability, and easier CE certification compared to bipedal models. The company employs a four-layer AI architecture, from fleet coordination to whole-body control, leveraging vision-language models for deterministic task execution. Humanoid AI focuses on deploying flexible robot fleets capable of diverse tasks, aiming for performance exceeding human speed through reinforcement learning. Despite less capital than US competitors, strategic partnerships with Bosch and Schaeffler, and a commercial-first approach, position them for rapid growth and European manufacturing. They are expected to secure more funding soon.

HMND 01’s four-layer brain starts at layer zero, just like European elevators. That’s oddly fitting, given that its maker, Humanoid AI, is based in a sleek tower high above downtown London. It’s not the sort of place you might expect to find the grease and gears and rigging of a contemporary robotics startup, but it’s here that Humanoid is quietly assembling one of Europe's most credible answers to the humanoid robot race.

While American rivals like Apptronik, Figure and 1X plus a wave of high-volume Chinese manufacturers like Agibot, Unitree and UBtech dominate the headlines, Europe has been building its own humanoid contenders. That includes Humanoid in the UK, Neura Robotics in Germany and a handful of others like Oversonic Robotics in Italy. Typically – though not with Neura Robotics – these European startups have far less capital than their U.S. counterparts … and therefore more urgency about shipping.

I spent an hour with Humanoid’s chief product officer Sotirios Stasinopoulos walking around the company’s UK HQ, talking about Humanoid’s strategy and timing.

Obviously Greek in origin, Stasinopoulos spent over a decade in China working for top robotics companies like UBtech as well as starting several companies of his own. That’s a great real-world education for building humanoid robots on a budget and a timeline, and he’s put his experience to good use quickly. Humanoid was founded in 2024 by Artem Sokolov, who scaled the Sokolov jewelry business into a billion-dollar manufacturing firm. Still just 130 people, Humanoid has already signed deployment deals with Schaeffler, Bosch and Siemens and launched both bipedal and wheeled humanoid prototypes, all without raising the billion-plus dollars of venture capital its best-known competitors enjoy.

The most visible bet Humanoid has made is one you can see the moment you walk the floor: the flagship humanoid robot it’s focusing most of its development effort on rolls rather than walking.

The company has two platforms, a wheeled mobile manipulator called the HMND 01, and a bipedal robot. But for now, roughly 90% of its engineering effort goes into the wheeled version, for reasons Stasinopoulos laid out.

"More than 85 or 90% of the use cases can be covered by a wheel-based platform," he told me, echoing an assessment I’ve heard from multiple robotics companies. Any wheeled platform is more energy efficient because it doesn’t burn power just standing upright, is more stable thanks to a very low center of gravity and therefore has a bigger working envelope.

"It can go from zero up to 1.5 meters fully extended, 1.5 meters away from the robot," he said, talking about HMND 01’s hands. The current Alpha version lifts 15 kilograms (33 pounds); the next will handle 20 (45 pounds).

But the truly decisive form factor decision is regulatory. Industrial customers typically need CE certification, and that’s a lot easier for a wheeled platform than bipedal.

"There’s no ISO at this stage that can cover bipedal technology for industrial frameworks, because at any time you may fall," Stasinopoulos says. "It means a lot of redundancy. The certification is still not mature. We’re in these committees … we believe maybe late next year or early '28 you might have something."

The wheeled platform, by contrast, can be certified under existing frameworks by borrowing from two of them at once: the autonomous mobile robot (AMR) standard and the collaborative-robot, or cobot, industrial framework.

"If we wanted to have a product with CE certification by next year, this was the way to go," he said.

Humanoid is targeting a fully certified product by late 2027.

The biped is still around, though. But it has a different target market.

"We are retargeting this to a more service and household application, but as a mid- to long-term strategy," Stasinopoulos said. Bipedal technology, he added, still isn’t robust enough for the 20-hours-a-day industrial duty cycle Humanoid designs around, and the home safety case isn’t yet completely solved. "You definitely don't want a big heavy metal thing falling on a pet or a baby."

Humanoid’s hardware philosophy is pragmatic. So is its software architecture: organizing robotic intelligence into four layers, numbered from three down to zero.

At the top sits System 3, an agentic fleet coordinator that ingests tasks from a customer’s warehouse management or ERP system and distributes them across robots based on their location, battery status and — because the end effectors are modular — their current capabilities.

(At some point those end effectors should be hot-swappable, so a robot could theoretically switch out fully 20+ degree of freedom robotic hands used for sensitive assembly, perhaps, to simpler and more robust grippers for logistics and shipping department duties, perhaps.)

Below it, System 2 is a reasoning layer built on off-the-shelf vision-language models like Google's Gemini, which breaks each task into a deterministic, checkable workflow. System 1 is Humanoid's proprietary vision-language-action model, which executes discrete actions like picking an object off a shelf. And System 0 is the whole-body controller that turns those commands into actual motion.

I asked if that flow, essentially, is how you take probabilistic technology like an LLM and use it as part of a physical AI software stack to power tasks that are deterministic.

"Exactly," Stasinopoulos said. "You ask the agent or the VLM to create this workflow for you, but you need to check it and make sure it's what you want to achieve."

"We're not trying to deploy single robots for single applications," he said. "If you only want a robot to do one thing, you might as well get an old-school automation solution."

The goal, instead, is a workforce as flexible as humans: robots that stock shelves in the morning, feed machines in the afternoon and reorganize inventory overnight.

Speed is always a question when you’re looking at robotic – and especially humanoid robotic – solutions to labor. How fast are they compared to our most common reference point: a human?

Humanoid says its robots are now performing core tasks at roughly 80% of human speed and success rate on some tasks, up from around 60%, and it expects to approach — and eventually exceed — 100%.

That’s already pretty impressive, given the speed I’ve seen on humanoids so far.

That last claim though – exceeding human speed – rests on reinforcement learning, not teleoperation-based human-gathered data.

"If we're only to use imitation learning, based on the data we collect with humans, then human performance is the limit," Stasinopoulos said. "Reinforcement learning does not have this limit."

The company now leaves robots working overnight, with System 2 automatically judging good and bad executions and feeding the wins back into a shared base model. "One robot will do something better, System 2 will say this is a good execution, retrain the base model, and then all the robots get the improved model," he said. "The progress is exponential."

That’s impressive, but not shocking given that the company has so quickly executed on launching significant humanoids. Two things explain the pace, he argued.

"We don’t develop technology and see how it can be used," he said. From his arrival in September 2024, the company chased big customers armed with little more than a presentation, then designed to their requirements. The second is what he called “second-mover advantage:” the team includes more than 50 veterans from Boston Dynamics, Sanctuary AI, 1X and other top robotics and tech companies.

"We’re making use of all the knowledge and all the mishaps that other companies had to pay millions or billions of dollars to learn," Stasinopoulos told me.

There’s a reason Stasinopoulos moved back to Europe from China: robotics isn’t just tech. It’s also geopolitics: nations and alliances with top-notch robots and humanoid robots will have labor and productivity advantages others will struggle to compete with.

And, in an era of supply chain uncertainty when it comes to magnets, rare earths, key technologies like actuators and more, everything here is political, too.

So the supply chain is also a strategic question, Stasinopoulos say, and Humanoid will manufacture in Europe for European customers as well as pursuing a "China plus one" approach in its supply chain to avoid being locked in if geopolitics turn. The most critical components come through strategic partnerships: Humanoid’s actuators are co-designed with Schaeffler, while compute runs on Nvidia.

"For the most crucial parts of the solution, we'll work with strategics," he said. "For the non-so-crucial components, we'll keep an open mind."

Actuators are key here: they’re commonly 50% of the cost and complexity of the robot, and efficiency, power, thermal management and longevity are all critical. That’s why the European partnerships Humanoid has forged with Bosch for manufacturing and Schaeffler for actuators are so important. Those strategic deals turn suppliers into partners and customers, with Schaeffler signing up for at least 1,000 of Humanoid’s robots, and Humanoid getting access to enough production capacity to build 100,000 humanoids by 2031.

There’s a lot you can do on ingenuity and hard work. But a little extra capital wouldn’t hurt either, and it sounds like that could be coming soon.

Humanoid has raised $74 million in funding so far, but local rivals like Neura Robotics in Germany have raised over $1 billion. To compete, Humanoid has to boost its war chest, and while Stasinopoulos couldn’t comment specifically, that seems to be in the cards fairly soon.

That would only strengthen Humanoid’s position as competition intensifies. Figure, Apptronik and other U.S. rivals have all raised billion-dollar war chests, and matching that scale will require significantly more capital.