Picture an operator on an assembly line, hands busy fitting the gears of a gearbox. There is no paper manual on the bench and no screen to glance away at. The next step — which component, which orientation, how it locks into place — appears directly in the operator’s field of view, projected by an augmented reality headset. A robot works alongside, fetching the parts and handing them to the operator. This is the scene at the heart of BI-REX’s pilot within MOTIVATE XR.
Now change one detail. What if the robot were a humanoid?
For most of industrial history, that question lived in science fiction. The robots that earn their place on factory floors look nothing like us: they are arms bolted to the ground, fast and precise within a cage or a carefully defined collaborative zone. The humanoid shape — two legs, two arms, a head full of sensors — has always been far harder to build and, above all, far harder to control. What has changed is not the metal. It is the intelligence that moves it. The field now calls this Physical AI, and it is the reason the humanoid question is suddenly worth asking out loud.
Human and Robot, Side by Side
BI-REX is one of the pilots in MOTIVATE XR, a project whose goal is to build a leading suite of XR tools for authoring, publishing and experiencing training and assistance in industrial settings. Our pilot tackles one concrete, demanding version of that challenge: human–robot collaboration in the assembly of a gearbox, with the instructions the operator needs projected through an augmented reality headset.
The value is easy to lose if you only look at the hardware. The real problem in modern manufacturing is rarely raw speed; it is complexity and change. Product variants multiply, batch sizes shrink, and the person on the line must hold more information in their head while their hands are occupied. Printed instructions sit on a bench, out of reach and out of sync with what is in front of the worker. An AR headset changes that relationship. The guidance follows the gaze: the right step, anchored to the right part, at the right moment, without the operator ever breaking focus. The result is lower cognitive load, fewer errors, safer handling, and the flexibility to switch between tasks without a lengthy relearning curve.
The robot in Fig. 1 is, today, a collaborative robot — a cobot — mounted on top of an AMR (autonomous mobile robot). It is precise and tireless, and it shares the workspace with the human instead of being fenced off from it. But it is still, fundamentally, a machine that does exactly what it is said, in exactly the space it has been given. It does not improvise. It does not pick up a tool it has never seen and works out what to do with it. That is the line Physical AI is starting to cross.
Fig. 1 — The BI-REX Mobot used in MOTIVATE XR Pilot 5
The Missing Piece: Physical AI
Most of the AI we have met so far lives on a screen. It reads, writes, answers questions, generates images. It is extraordinarily capable, and it is also disembodied: it never has to keep its balance, judge how hard to grip a bolt, or recover when a part slips. Physical AI is the branch that does. It is artificial intelligence that perceives and acts in the real world — reasoning about space, objects, forces and movement, and then controlling a body that must obey the laws of physics in real time.
This is the piece that kept humanoids stuck in research labs for decades. A humanoid body is brutally hard to coordinate. Every joint must move in concert, balance must be maintained through every step and reach and contact with the world must be gentle enough not to break things yet firm enough to get work done — all of it continuously, not as a scripted sequence. Writing that behaviour by hand, movement by movement, simply does not scale.
So why bother with the human shape at all, when a wheeled arm is easier? Because our world is already built for us. Factories, workstations, tools, stairs, handles and doorways are all designed around the human body. A machine shaped like a person can, in principle, step into those spaces and use those tools without anyone rebuilding the factory around it. Humanoid form is not a vanity project; it is a bet on compatibility with the world we already have.
A Brain for the Body: Cortex and Cerebellum
Here is the idea that makes Physical AI click for most people. Your own ability to move well does not come from one monolithic “mind.” It comes from a division of labour inside the brain.
Fig. 2 — Physical AI is splitting a humanoid’s “mind” the way our own brain does: a reasoning part that decides what to do, and a motor part that works out how to move.
The cerebral cortex handles the deliberate part of movement — understanding the goal, planning a sequence, drawing on language and knowledge to decide what to do. The cerebellum handles a different job entirely: coordination, balance, timing, and the fine-tuning that makes a movement smooth rather than jerky. The cortex decides that you will pick up the cup; the cerebellum makes sure your hand arrives there without overshooting and closes around it with just the right force. Damage the cerebellum and you can still want to move and still know how you intend to — but the execution falls apart.
Physical AI for humanoids is converging on exactly this split, and two models from NVIDIA map onto it almost perfectly.
The “cortex” is Isaac GR00T, an open foundation model for humanoid reasoning and skills. It is what the field calls a vision-language-action model: it takes in what the robot sees together with an instruction given in plain language, reasons about the scene, and decides what the robot should do. Fittingly for a brain analogy, GR00T is itself built on the principle of thinking “fast and slow” — a slow, deliberate part that plans, and a faster part that turns plans into action. It is the part of the system that understands the task.
The “cerebellum” is SONIC, a foundation model for whole-body control. Instead of being hand-programmed move by move, SONIC learns from an enormous library of human motion — well over a hundred million frames, representing hundreds of hours of movement — so that a single model can walk, crawl, keep its balance, reach and manipulate objects. Where older robots needed a separate, painstakingly tuned controller for each new skill, SONIC treats movement itself as one thing to be learned at scale. In testing on a real humanoid, a single SONIC model reproduced well over a hundred different motions with near-perfect reliability. It is the part of the system that knows how to move.
Put the two together and something new appears: a machine you can tell, in ordinary language, what you need — and that can then work out how to move its whole body to deliver it. When researchers connected a reasoning model like GR00T to SONIC’s whole-body control, the pair handled mobile manipulation tasks — moving through a space and acting on it — with a success rate in the mid-nineties.
A note of honesty belongs here. These are research results and controlled demonstrations, not a humanoid that can walk onto any assembly line tomorrow and start building gearboxes. The gap between an impressive demo and a dependable colleague is wide, and it is mostly made of unglamorous things: reliability, safety, repeatability, cost. But the direction of travel is unmistakable, and it is moving faster than almost anyone predicted even two years ago.
Where XR Comes In
Physical AI gives a machine a brain and a body. It does not, on its own, give the person working next to it a way to tell the machine what is needed, or to understand what it is about to do. That is the gap XR is well placed to fill — and XR is the technology MOTIVATE XR is built around. In our pilot we see three concrete links.
The first is the task itself. Every assembly step in the pilot is already authored as structured, spatial information: which part, where it sits, how it is oriented, in what order. Today that description is rendered as holographic guidance for the operator. But it is also exactly the kind of input an intelligent robot need — a goal anchored to real objects, rather than a list of hard-coded coordinates. Author the step once, and who carries it out — the operator, the cobot, or one day a humanoid — becomes a question of assignment, not of reprogramming.
The second is teaching. SONIC can be teleoperated by a person wearing a VR headset, either with full-body tracking or with just a few key points such as the head and hands: the operator moves, and the humanoid follows in real time. The demonstrations recorded this way can then be used to train vision-language-action models such as GR00T. Instead of writing code, an expert shows the machine how a job is done — in the same kind of immersive environment already used to train people.
The third is trust. A humanoid working at arm’s length is only acceptable if the operator can anticipate it. Through the headset, the robot’s intentions can be made visible: the part it is about to pick, the path its arm will follow, the zone it is about to enter. This is what turns a machine that merely shares the workspace into a teammate the operator can rely on — and it is where the AR headset in our pilot stops being just a better manual.
MOTIVATE XR – Pilot 5 – Human Robot Hybrid Manufacturing Demo
Link — The BI-REX pilot in action: AR-guided human–robot collaboration on gearbox assembly.
What It Means for the Worker
It is worth being clear about the point of all this, because “humanoid robots in the factory” can sound like a story about replacing people. The premise of MOTIVATE XR, and of Industry 5.0 more broadly, is the opposite: technology in service of the human, not instead of them. A humanoid teammate earns its place by taking on the parts of a job that are physically punishing, monotonous, or awkward — the lifting, the holding, the hundredth identical repetition — while the person keeps what people are still far better at: judgement, dexterity guided by context, and solving the problem nobody anticipated.
The deeper lesson from our pilot is not about any robot at all. It is that the value lives in the shared understanding — in the operator and the machine working from the same information, in the same space, at the same moment. Physical AI is making machines better at listening and acting. XR is making information something a human and a machine can genuinely share. Those two trends are heading toward the same place.
Conclusion
Return to where we began: an operator fitting a gearbox, guided by an AR headset, a robot at their side. Today that robot is a cobot and the headset guides the person. It is not hard to imagine a version a few years out where the teammate is a humanoid — carrying a cortex that understands the task and a cerebellum that knows how to move — and where the same XR layer guides both of them from a single, shared set of instructions.
BI-REX’s pilot is a concrete step on that road. The gearbox is small and specific; the question underneath it is not. How do people and increasingly capable machines share a task well? MOTIVATE XR exists to answer that question with tools rather than speculation, and our pilot is one place where the answer is being worked out, one assembly at a time.
Author
BI-REX Competence Center
Roberto Giudiceandrea is an Automation and Robotics Application Engineer at BI-REX Competence Center. Roberto focuses on robotics integration, virtual commissioning, and enabling technologies for Industry 4.0, with expertise in digital twins and collaborative automation solutions.
References
[1] NVIDIA Isaac GR00T N1 — open foundation model for humanoid robots (GTC 2025), coverage: https://www.techradar.com/pro/the-age-of-generalist-robotics-is-here-nvidias-latest-groot-ai-model-just-took-us-another-step-closer-to-fully-humanoid-robots
[2] NVIDIA GEAR-SONIC — whole-body humanoid control, model card: https://huggingface.co/nvidia/GEAR-SONIC
[3] “Nvidia’s SONIC teaches humanoids to move,” AI Business: https://aibusiness.com/robotics/nvidia-s-sonic-teaches-humanoids-move
[4] MOTIVATE XR project: https://motivatexr.eu/
[5] MOTIVATE XR within the XR5.0 ecosystem: https://xr50.eu/synergies/motivate-xr/
