The most consequential question in humanoid robotics is not which company builds the best body. It is whether the body and the brain will ever be sold separately. Today almost every commercial humanoid arrives as one indivisible product: mechanics, joint controllers, perception, motion policy and task software from a single vendor, under a single warranty, with a single phone number to call. The personal computer looked like that too, until 1981. A genuinely software-defined humanoid, one where the intelligence layer is bought, replaced and upgraded independently of the mechanism, would change who captures the value in this sector and who carries the risk.
For anyone evaluating humanoid robots for European deployment, this is not a philosophical question. It decides whether the machine you sign for in 2027 can be re-skilled in 2030 by someone other than the company that built it, what happens to your support contract if you try, and whether your CE marking survives the attempt.
What the PC analogy actually claims
The personal computer did not become an open platform because anyone wished it. Three specific things happened, and the same three are the test for robots. IBM published the bus specification and a technical reference, so outsiders could build cards that worked. The operating system was licensed separately from the hardware, so MS-DOS ran on machines IBM did not manufacture. And in 1982 Compaq legally reimplemented the BIOS, proving the interface could be crossed without the platform owner’s permission. Documented interface, independently licensed software layer, demonstrated crossing. Miss one and you get a console, not a PC.
Applied to humanoids, the score in 2026 is roughly one and a half out of three. The software layer is starting to be licensed separately. The interface is not documented in any portable way. And nobody has yet put a third party brain on a certified commercial humanoid, in production, against the body vendor’s wishes.
The split has already started, at exactly one layer
Google DeepMind’s work is the clearest evidence that the intelligence layer is becoming a product in its own right. Gemini Robotics 2, announced on 31 July 2026, is presented not as a robot but as a layer: a vision-language-action model that turns camera and language input into motor control, now extended to whole-body control across legs, torso, arms and fingers. DeepMind states the model can be adapted to a new bi-arm platform in a matter of hours, and that the on-device variant can move to new hardware on fewer than 200 demonstrations collected in a single session. The named partners, Apptronik, Boston Dynamics and Agile Robots, are three unrelated bodies sitting under one model family. That is the definition of the layer separating.
NVIDIA pushed further down the stack. On 31 May 2026 it announced the Isaac GR00T Reference Humanoid, a Unitree H2 Plus body with Sharpa Wave hands, Jetson Thor compute and the Isaac GR00T software platform, pitched as giving research teams advanced hardware and an open software stack without requiring proprietary platforms. Read that carefully before repeating it. The announcement publishes no CAD, no bill of materials and no schematics. “Open” here means an open software stack running on a documented commercial body, which is a real step and is not open source hardware.
Physical Intelligence took the opposite end, publishing the weights and training code for its generalist policy, so a model trained on somebody else’s fleet can be inspected and fine-tuned by a team that owns none of it. And in December 2025 Unitree opened a developer platform where owners share action sequences and training datasets between units, the closest thing the sector has to an app store. Its catalogue today is largely pre-programmed motion, dance and martial arts routines, not manipulation policies carrying a safety case. That gap is the entire remaining problem.
Where a software-defined humanoid is actually possible
A humanoid is not two things, a body and a brain. It is six layers, and only the bottom three of this list are open in any meaningful sense today.
| Layer | What it does | Replaceable in 2026? |
|---|---|---|
| Mechanics and actuation | Joints, hands, structure, power | No. This is the product you bought. |
| Low-level control and safety | Torque loops, limits, emergency stop | No. Vendor firmware, certified with the machine. |
| Whole-body control | Balance, gait, reach coordination | Rarely. Occasionally exposed as a constrained API. |
| Policy layer (VLA) | Perception to action for a given task | Yes. This is the layer that opened. |
| Orchestration and fleet | Task assignment, monitoring, logging | Yes. Mostly ordinary enterprise software. |
| Application and integration | WMS, ERP, workflow, reporting | Yes. Already a normal integration project. |
The dividing line is not technical difficulty. It is where safety liability sits. Everything below the line is commercial software and can behave like commercial software. Everything above it is a certified machine, and certified machines do not accept arbitrary code.
For buyers: when a vendor tells you the platform is open, ask which of those six rows they mean. In nearly every conversation in 2026 the honest answer is the bottom three, and the interesting question is what it takes to reach the fourth. This maps directly onto what a humanoid robot actually runs as an operating system, where the same layering shows up as middleware, runtime and application.
What is still missing before the layers separate
Four things stand between today’s arrangement and a market where you buy mechanics from one company and intelligence from another.
There is no bus. Humanoids have no agreed schema for observation spaces, action spaces, joint conventions, latency guarantees or sensor timing. Every port of a model to a new body is bespoke engineering, which is precisely why “a few hours” is impressive enough to put in a press release.
Models port, skills do not. A policy trained on a three-finger gripper does not transfer to a five-finger tendon-driven hand, because the action space differs and so does the contact physics. The architecture is portable. The competence has to be rebuilt with new demonstrations, which is time, teleoperation labour and money. Anyone selling brain portability should be asked what the retraining actually costs, a question the current state of VLA models in robotics makes concrete.
Nobody certifies a brain. Conformity assessment in Europe applies to a machine placed on the market, not to a model published on a repository. There is no route today by which a policy vendor can hand you a certificate that transfers across bodies.
Support collapses at the boundary. The moment third party software commands a joint, the incident investigation has two defendants and no agreed method of separating them. That is a commercial problem long before it is a technical one, and it is the reason vendors resist opening the third row of the table. The open source projects worth following in humanoid robotics run into the same wall at exactly the same layer.
The counter-case: integration is winning right now
Vertical integration is not a failure of imagination. Figure builds its own hardware and its own Helix vision-language-action model, having moved off an external language provider, and treats full stack ownership as the strategy rather than a stage to grow out of. The engineering arguments are real: latency budgets are easier to hold inside one codebase, the data flywheel only closes if you own both the robot collecting the data and the model consuming it, and when a robot drops a crate there is exactly one accountable party.
The honest position in 2026 is that vertical integration produces better robots today and modularity produces better markets later. Nobody credible can tell you when later begins.
The PC analogy also has a survivorship problem. Smartphones did not decompose. Cars have not decomposed. In both cases the safety case, the liability structure and the software update chain kept the stack welded together, and the “open” layer stopped at the application store, curated by the hardware owner. That is at least as plausible an outcome for humanoids as the PC path.
European regulation may decide this before engineering does
Swapping robot software is not a neutral act in the EU, and this is where the abstract argument becomes a purchase decision. The Machinery Regulation (EU) 2023/1230 applies from 20 January 2027, and its provisions on substantial modification say that an operator who substantially modifies a machine becomes its manufacturer. A modification counts as substantial when it was not foreseen by the original manufacturer, happens after the machine is placed on the market, and creates a new hazard or increases an existing one. A software change can meet that test. Updates the manufacturer documented and anticipated generally do not.
Annex I now also covers safety components with fully or partially self-evolving behaviour based on machine learning, which pulls a notified body into the loop for exactly the kind of learning policy this whole article is about.
For buyers: the practical consequence is that hardware and software separation in Europe will not arrive as a free-for-all. It will arrive as vendor-approved catalogues, where the body maker validates what may run, publishes it as foreseen use, and takes a margin on the listing. Closer to a console than to a PC. Anything outside that catalogue puts the technical file, the declaration of conformity and the CE marking on your desk, not the vendor’s, which is the same trap described in how the EU Machinery Regulation treats humanoids as a new category.
What to do about it in 2026
Do not buy on a prediction about market structure. Buy on clauses, because the clauses are enforceable and the prediction is not. Five are worth insisting on before signature.
- Interface documentation. Written specification of what the robot exposes: which control modes, which telemetry, what latency, and whether the whole-body layer is reachable at all.
- Data and model portability. Who owns the demonstrations your staff collect during deployment, in what format you can export them, and whether a policy fine-tuned on your data leaves with you.
- Approved software list and a route onto it. If the vendor operates a catalogue, get the admission process in writing, with timescales, so a supplier you choose can be validated rather than refused.
- Warranty behaviour under third party software. Not whether it voids the warranty, everyone says it might, but which specific subsystems remain covered and how an incident is apportioned.
- Exit. What still functions when you stop paying the software subscription, and whether the robot degrades to a safe manual mode or to scrap. The answer belongs in the contract, alongside the update and rollback process for deployed fleets.
The layers will separate, because that is what happens to every hardware market once the software becomes the harder problem, and here it already is. But the separation will be slower and far more curated than the PC comparison suggests, and the buyers protected in the meantime will be the ones who wrote it into a contract in 2026 rather than the ones who waited for a standard nobody has drafted.
Frequently asked questions
Can I run Gemini Robotics on any humanoid robot?
No. As of August 2026 Gemini Robotics 2 and the on-device variant are available to selected early access partners rather than as a general product, and the embodied reasoning model is reachable through Google AI Studio. Even with access, moving the model to a new body requires integration work and fresh demonstrations, so it is a porting project rather than an installation.
Will there be a Windows for humanoid robots?
Not in that shape. ROS 2 is middleware rather than a consumer style operating system, and it does not define the policy layer where the value now sits. The closer analogy is a foundation model licensed separately from the body, running inside a vendor runtime that stays certified with the machine.
Does changing robot software void CE marking in the EU?
It can. Under the Machinery Regulation (EU) 2023/1230, applying from 20 January 2027, an operator who makes a substantial modification becomes the manufacturer of the machine. A software change qualifies where it was not foreseen by the manufacturer and creates a new hazard or increases an existing risk. Updates the manufacturer documented and anticipated normally do not.
Should I delay buying a humanoid until the market opens up?
No. The payback case for a humanoid rests on a specific task, a specific shift pattern and a specific labour cost, not on future market structure. The defensible move is to buy for the task and protect optionality contractually, through interface documentation, data portability and exit terms.