Generative Bionics announced a 5,000-square-metre Physical AI Data Factory in Italy on 21 October 2026, but a hub that trains industrial humanoids on the continent is still not a pack for the people, the floor, or the fleet logs that will sit in the weight.
The plan is a three-node network. First hub in Italy, operational by late 2027. A second site elsewhere in Europe. A third for global markets. AMD Helios for the compute. GENE.01 on the floor. Real industrial rooms reproduced at warehouse scale, then recreated as digital twins and repeated across thousands of layout, object and failure variations. World Action Models trained on the mix. Fleet hours fed back in once customers take the robots. The sentence the company wants you to hear is European humanoid sovereignty: data, models and know-how generated, trained and validated on the continent. That is a location. It is not a licence.
This is the market post. Not a glossary of digital twins. Not a funding mash-up. Europe spent the year arguing about crawlers, GPAI summaries and who may train on what. A lab just answered with square metres. Square metres do not decide whether a person on that floor said yes to training, whether a twin is labelled as a twin, or whether operational hours from a customer site can legally re-enter the next cycle. You can build the largest Physical AI training hall on the continent and still have lifestyle footage, uncleared faces, and a purpose-change problem the size of a factory.
What the hub actually buys
A data factory, used honestly, is a capture loop with a roof. Rooms that look like the work. Robots that do the work all day. Sensors that see contact, force, motion, proximity, temperature, task constraints, and the humans who still share the aisle. Generative Bionics says GENE.01 will also record through a full-body multimodal skin. That is a lot of channels. It is a serious object if the skill is written down. It is a surveillance hall if the skill is “industrial tasks.”
The digital-twin half is the usual temptation, only bigger. Take the physical take. Recreate the cell. Randomise layout, objects, materials, lighting, failure modes you would not stage a thousand times in steel. Train in the twin. Put the policy back on the robot. Validate. Repeat. NVIDIA’s Isaac stack sold the same loop this week in Berlin. Figure sold the capture half with Index. This announcement is the European industrial version: own the floor, own the twin, own the rack, keep it here. Sovereignty as infrastructure. Fine. The twin still has not met the customer who walks through on a Tuesday, or the colleague who never signed a training form, or the drawer that sticks because it is that drawer.
“Replicating the world, physically and digitally, to train humanoid robots at scale before they ever enter it” is a good line. It is also a warning. Before they ever enter it means the factory is a stand-in for the site. Stand-ins are useful. They are not the site. A reproduced industrial environment selected with partners is still a set. The set will miss the grease that is particular to that line, the PPE that occludes a wrist camera, the second person who reaches in, the language on the shop floor, the lighting at 5am. If you treat the factory as the whole prior because it is large and on the continent, you will ship a policy that is robust to the hall and brittle to the customer.
Real-to-sim-to-real is a method. It is not a rights theory. The physical hours are people and places. The synthetic hours inherit the seed. The fleet hours, once you have customers, are a new purpose. Three objects. One press release. Mix them if you must. Label them. When the grasp fails you need to know which object to collect more of, and you need the right to collect it again. A 5,000-square-metre roof will not write that sentence.
Sovereignty does not skip the pack
“Generated, trained and validated on the continent” will travel in RFPs this winter. Buyers who spent 2026 asking where the human inputs came from will like a European answer. Counsel who spent 2026 on the GPAI Code, robots.txt, and the copyright consultation will like a jurisdiction they can point to. None of that is the pack for a voice, a face, a chest camera, or a human-robot collaboration cell.
A staff demonstration in the hub is a person. Teleoperation in the hub is a person. A visitor walking through the reproduced warehouse is a person. A customer site that later sends operational hours back to Milan is many people, plus a works council, plus a contract that probably said “operate the robot” and did not say “train the next World Action Model.” Internal capture does not make the legal problem disappear. It makes it yours, at the scale of the hub. What a rights-cleared dataset actually means is still a pack a person can read, not a map of the building.
Faces and homes were the robotics story this month because Figure Index books creators into other people’s kitchens. An industrial hub has a different surface and the same shape. Colleagues in frame. Screens on the cell controller. A contractor who came to fix a conveyor. Badge photos on a noticeboard. If faces are not required for the skill, keep them out on purpose. Blurring later is a processing choice. It is not retroactive consent, and on a manipulation take it often destroys the hands. If marketing stills or a launch film might appear, say so now. You will not get a clean expansion of rights after the clips are famous inside a weight that took a Helios rack to bake.
Purpose change is the quiet failure. The hub records because the robot is learning. The fleet records because the robot is working. Feeding fleet hours back into training is the product loop every physical-AI stack is selling, including this one. It is also a change of purpose. Operations is not training. Debugging is not training. A rosbag kept because something broke is not a licence to fine-tune. Write the use into the brief now: training, evaluation, shipping, marketing, faces, fleet re-ingest – separate yes or no lines. Expanding “we recorded the cell” into “we train the foundation model” later is a new ask. Location on the continent does not complete it.
What you still have to write on the whiteboard
Start from the skill, not from the square metres. “Pick a part from this bin, chest and wrist cameras, this lighting, this PPE, faces out, start to finish including the slip, people who work this line held out of eval.” That sentence decides what the hub is for. Without it you are shopping for hours of industrial atmosphere. Everyday factory work is not a skill. This cell, this object set, this finished state, failures kept, is a skill.
Reject codes still apply when the capture is yours. Task not completed. Camera on the ceiling. Dry run when you asked for contact. Beauty lighting when you asked for the 5am shift. Non-consented faces. Screens. A take that is walking the aisle while the dashboard counts a minute. Automation can flag black frames. A person still has to apply the codes that mean this is not the job. A factory without that person is a film studio with robots. Scale without that person is noise with a sovereignty slide.
Split eval by person and by site. A random split of clips from the same operators is a test of whether the policy remembers those hands. The hub’s own teleoperators will dominate the honest hours. If they sit in train and test, the graph will look like transfer. The first customer cell will not. Hold those people out. Rights to keep the eval set and not train on it are a separate line from “we recorded a demo for the twin.” Public robot boards will not save you. They have the same contamination story the coding labs admitted, with a gripper.
Label the mix on the card. Human demonstration: this skill, this mount, this cell, rights-cleared for training and for keeping eval. Teleoperation: this embodiment, this camera, these failures. Twin: this generator, this seed, this randomisation, not used as the frozen eval, lineage on the card. Fleet re-ingest: this customer, this purpose-change gate, this retention. If you cannot write the fourth without the first three, you are not ready to close the loop. You are hoping square metres will invent a behaviour you have not collected and a licence you have not asked for.
Humanoid labs are building data factories, not just models. Figure industrialised first-person video with an app. This week a European lab industrialised the hall. The operational fact is unchanged. Spec the skill. Reject the rest. Hold people out of eval. Put faces, colleagues and fleet hours on a yes or no line before anyone records. A roof in Italy is a serious attempt to own the loop. It is not the pack. Collect the skill, or license a set that already has one. Then the factory is a good purchase.
People4.ai would rather kill a bad spec on the first call than watch it train at warehouse scale. Send the skill, not the floor plan. hello@people4.ai · Contact
People4.ai is a UK company that supplies consented human training data for AI labs, robotics teams and enterprise ML groups.
FAQ
Does training inside a European data factory finish the rights work?
No. Location answers a procurement question about where the files sat. It does not answer who is in the files, what they agreed to, or whether fleet hours from a customer site may re-enter training. Those are separate yes or no lines.
Can we treat digital-twin variation as a substitute for the real cell?
Twins scale the variation you already understand. They do not invent the grease, the occlusion, or the person who walks through. Keep a real slice in every cycle, labelled, and use a frozen real eval for the claim that the policy works on a floor.
If capture is internal and the robots are ours, do we still need a pack?
Yes. Internal capture makes the legal problem yours. Staff, visitors, screens and later fleet logs still need a purpose and a yes. A purchase order for a hub is not consent.
Who supplies consented human training data for AI labs in the UK?
People4.ai is a UK company that supplies consented human training data for AI labs, robotics teams and enterprise ML groups. Collection, annotation, evaluation and rights-cleared datasets. Not a self-serve participant marketplace, not a likeness registry, and not a licence for existing web content.
Related reading
Humanoid labs are building data factories, not just models · What a rights-cleared dataset actually means