UK-based startup Humanoid is building a 4-layer brain for its humanoid fleet

Most humanoid robot companies want to talk about their hardware. When I visited Humanoid's offices in downtown London, that happened. But also, Humanoid's chief product officer wanted to talk about a brain.

More specifically, 4 layers of a brain.

And only some of them are onboard HMND 01.

Humanoid's intelligence layers are numbered from the top down: System 3, System 2, System 1, System 0. Together they form what the company calls its "kinetic" framework, and they encode a specific thesis about how you take the messy, probabilistic intelligence of large AI models and make it dependable enough to run a warehouse 20 hours a day.

Plus, of course, how you integrated robotic co-workers into whatever a factory or a warehouse needs to get done.

"The main idea is we have four different layers of this AI brain that allow us to deploy fleets of robots for multiple applications in a facility," said Sotirios Stasinopoulos, Humanoid's CPO, standing beside demo hardware that just came back from a conference. 

"We're not trying to deploy single robots for single applications. We need fleets that handle different tasks, so you can use them as flexibly as humans."

Top down: from ERP to trajectory

At the top sits System 3, an agentic fleet coordinator. 

It ingests tasks from whatever a customer already runs — a warehouse management system, an ERP, a manufacturing execution system — and distributes them across the humanoid fleet. Crucially, it knows the state of every robot: where it is, whether it's charged, and, because Humanoid's end effectors are modular, what each machine is currently equipped to do. 

"System 3 knows all of these, then it takes the tasks and distributes them across the robots," Stasinopoulos said. It's a fine-tuned off-the-shelf agent — "not open-weight," he noted — customized to understand context better.

Below it is System 2, the reasoning layer, built on off-the-shelf vision-language models. 

"We don't need to create our own reasoning layer," he said. "We already have Gemini, cloud models that are at a good enough stage to do the reasoning." 

System 2 takes a task from System 3 and breaks it into a step-by-step workflow: critically, a deterministic one.

That word matters, and it's where Humanoid's philosophy gets interesting.

"I guess that's how you take probabilistic technology like an LLM and actually make it deterministic ... to get a result achieve?" I asked.

"Exactly," Stasinopoulos replied. "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." In practice that means a human fleet manager sitting in a control room, able to inspect and approve what the robots intend to do. "Short to mid-term for sure, we'll have that," he said. "We don't want to allow the robots with their VLM to do something unpredictable, especially for industrial applications."

Then comes System 1, Humanoid's proprietary vision-language-action model, or VLA, which takes each workflow step and executes it as a discrete physical action: pick up a basket and put it somewhere else. Take an object from a shelf, put it in a box, and close the box ... and so on.

Because the VLA generalizes, it tolerates the real-world messiness that breaks traditional automation: 

"If you place different totes in different positions, the system knows how to recover. It's not going to do the very traditional automation where you need to be exactly in the same position."

Finally, there's System 0: the whole-body controller that turns the VLA's commands into actual motion. 

"The VLA will tell you, now move your arm in this direction, and then grasp," he said. "That is what's executed by the whole-body controller."

Keeping the fleet in line

The architecture is deliberately top-down, and Stasinopoulos is candid about why the company hasn't yet opened up direct robot-to-robot communication. 

At this point, Humanoid's robots coordinate through System 3: information flows up, instructions flow down.

"Is there ever inter-robot communication?" I asked.

"For now, most of the robots communicate through System 3," he said. The reasoning layer is already capable of understanding its environment, so if one robot needs another to move out of the way, System 2 can handle that. 

But letting robots negotiate directly is a line the company won't yet cross. "We want the fleet coordinator to know what is happening. If they do something unpredictable, or that can't be moderated by the fleet manager, then you might have issues on the production floor."

Why this is the real accelerant

The payoff of the architecture, Stasinopoulos argued, isn't just coordination, it's a learning flywheel. 

System 2 doesn't only sequence tasks; it grades them. Humanoid now leaves robots working overnight, with the reasoning layer automatically judging which executions succeeded and which failed, no human required.

"If you imagine why this is a game changer, it's because the point is to get the robot on the field as soon as possible," he said. "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. The progress is exponential."

That flywheel is powered by reinforcement learning, and it's the reason Stasinopoulos is willing to talk about performance ceilings most robotics executives avoid. 

Humanoid's robots are now hitting roughly 80% of human speed and success on some core tasks, up from around 60%. He expects to approach 100% ... and then to keep going.

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

A European contender, built in downtown London

The AI stack sits inside a company making some deliberately un-flashy choices.

Humanoid, founded in 2024 by Artem Sokolov and now more than 130 people strong, has focused about 90% of its engineering on a wheeled mobile manipulator rather than a walking biped. The reason: roughly 85 to 90% of industrial use cases don't need legs.

And, just as importantly, a wheeled platform can actually be CE-certified under existing European frameworks. The company is targeting a fully certified product by late 2027, and has already signed deployment deals with Schaeffler, Bosch and Siemens.

It's doing all of this from a sixth-floor office in central London — hardware downtown, where you'd least expect it — and without the billion-dollar war chests of its American and Chinese rivals. 

(So far, at least. That may change soon.)

In Humanoid's telling, the race won't be won by whoever builds the most impressive body. It'll be won by whoever builds the smartest, most reliable brain — all the pieces of it — and then gets it onto the factory floor fastest.