The interface is being written now, and it does not say humanoid

On 5 November 2025 SAP announced a physical AI program that is much larger than the coverage suggested. Eleven robot makers are named: Aimbo, AgiBot, ANYbotics, Apptronik, Booster Robotics, Boston Dynamics, Galbot, Humanoid, NEURA Robotics, RobotEra and Unitree. So are the customers: Bosch, Vodafone Procure and Connect in Duisburg, Mahindra and Mahindra, Martur Fompak, Tetra Pak, Arçelik-LG, BITZER, Sartorius, and SAP's own warehouse. So are eleven integration partners, Accenture and Capgemini among them.

The work is real rather than theatrical. At BITZER, NEURA's 4NE1 took picking tasks from SAP S/4HANA for Extended Warehouse Management and stayed in sync with the business logic while doing them. In a proof of concept run through January and February 2026 with Martur Fompak, the Humanoid company's HMND 01 took warehouse tasks over the internet and wrote findings back into SAP in real time. At the Duisburg pilot shown at Hannover Messe in April, robots did visual inspection for damaged product, assessed pallet stacking and weight distribution, and flagged obstacles.

SAP describes Joule, its agent layer, as an interface for embodied AI, and its own architect frames the value as letting robots "know why, when and how to act" by grounding them in trusted SAP data. That is a claim on this seam, made by the company already installed at the customer.

But read the mechanism rather than the announcement. SAP's own description of the Bosch work has context-enriched EWM tasks routed through a partner's fleet management system. SAP is not talking to the robot. Something else sits in between, and that something is being built by a system integrator, per customer, per robot brand. Eleven robot makers and eleven integration partners is not a standard. It is a matrix, and somebody is hand-wiring every cell of it.

Underneath, the abstraction is further along than most people assume. NVIDIA's Isaac GR00T N1.7 is an openly licensed vision-language-action model described as cross-embodiment, meaning one policy is meant to drive different bodies, with Jetson Thor as the runtime. ROS 2 has carried a middleware abstraction for years in rmw, which lets vendors swap the underlying transport without touching application code, and the working parts of the modern stack, MoveIt, Nav2, LeRobot and GR00T itself, assume it.

The sharpest evidence that this layer is where value is moving is a company that changed sides. On 17 June 2026 Sanctuary AI announced it would stop waiting for humanoid hardware to reach mass commercialization and deploy its Physical AI on existing commercial robot platforms instead. It reported a plug insertion task on a moving conveyor at better than 99.5 percent success and a 2.54 second cycle time, matched to a Tier 1 automotive supplier's live line rate. A company that built one of the better-known humanoids concluded that the software which runs on anybody's body is the more valuable asset than the body.

So the bottom of the stack is being commoditized on purpose, by the largest chip company in the world, under a permissive license. Nobody should plan to win there.

Between those two there is a layer with a mature, boring, extremely instructive precedent. VDA 5050 is the interface the German automotive and mechanical engineering associations wrote so that automated guided vehicles and autonomous mobile robots from different manufacturers could take orders from one master control. It is MQTT carrying JSON, and it defines six channels: orders, instant actions, state, visualization, connection, and a factsheet.

That last one deserves attention from anyone thinking about routable capacity. The factsheet, added in version 2.0, is a machine-readable description of what a given machine can do. A fleet controller does not need to know the vendor. It reads the capability declaration and dispatches accordingly. That is the primitive underneath every routing argument this site has ever made, already specified, already deployed, and nobody had to invent a new philosophy to get it.

The standard is alive and moving. Version 3.0.0 was published in April 2026, developed by the VDMA materials handling association and the VDA under technical supervision from the Karlsruhe Institute of Technology, after a public consultation on GitHub that ran to early February. It adds a zone concept so robots with real autonomy can plan their own routes inside rule-bound areas, path sharing so central control can still manage traffic, and it stays backward compatible. Volkswagen Commercial Vehicles is reported to run the largest fleet on it.

Note carefully what VDA 5050 refuses to do. It does not plan routes, resolve traffic or assign jobs; that intelligence belongs to the master control. Functional safety is explicitly out of scope. And, in the words of one plain-language explainer of the standard, it is an interface and not a brain, and it does not handle vertical integration with production systems, which remains a separate implementation project for every organization that attempts it.

Read that last sentence twice. Even in the category where the abstraction layer exists, is mature, is open source and has an anchor customer the size of Volkswagen, the connection up into the system of record is still bespoke. The two halves have never been joined.

The joining is now being attempted, and it is being attempted at ISO. ISO 21423 specifies communication protocols for interoperability between industrial autonomous mobile robot systems from different vendors, and its scope covers the mobile robots, the fleet manager, and the enterprise resources that communicate with them. It sits at Draft International Standard stage in a ballot inside ISO Technical Committee 299, was developed with close to 90 participants, aligns on MQTT and JSON, and is expected to publish later this year. It is also notable for departing from a purely centralized model in favor of federated orchestration, where multiple fleet managers coexist.

That is not an adjacent development. Enterprise resources that communicate with the robots is the seam, named, in a standard that is weeks or months from publication.

The commercial version already exists too. InOrbit runs an orchestration platform with more than 30 robot brands in its ecosystem, and at Automate 2026 in June it ran a live multi-vendor demonstration with Ati Robotics, Kärcher, NEURA, Omron, Peer Robotics, Quasi Robotics and Unitree, the last of which brought quadruped and humanoid platforms. Its Business Execution System is described as translating orders from warehouse management, ERP and manufacturing execution systems into robotic execution. That is the founder's hypothesis, built, demonstrated, with humanoids in the room.

Two other efforts are worth knowing and neither is winning. The MassRobotics interoperability standard sits at the monitoring layer, where robots report status to other robots, fleet managers or ERP and warehouse systems, so it informs rather than commands, and it remains largely a United States effort without major deployments. Open-RMF, from the Open Source Robotics Foundation, orchestrates mixed fleets against building infrastructure and has seen minimal industrial adoption. In Asia no standard dominates, and Chinese vendors show interest in VDA 5050 without much sign of adopting it.

Here is the finding, and it is narrow enough to act on. VDA 5050 version 3.0 does not mention humanoids or mobile manipulators. ISO 21423 is written for industrial mobile robots. MassRobotics covers track-guided and autonomous transport robots. Open-RMF targets service robots. The entire middle layer, the one that connects a fleet to the system of record, has been specified for machines that move things from A to B.

A humanoid is not that. It is a mobile manipulator whose value is the manipulation, and the vocabulary that would describe its work, a task with a tool, a tolerance, a quality outcome, an inspection result, a qualification it holds, does not appear in an interface designed to dispatch transport orders. The factsheet concept is the right shape and the wrong contents.

Safety is on a similar trajectory and slightly ahead. ISO 10218 was overhauled in 2025, its first major revision since 2011, split into a part for manufacturers and a part for integrators. The old version excluded robots on mobile platforms outright, which took humanoids out entirely; the new one excludes only the mobility aspects, which lets the rest of a humanoid back in. Dynamically stable legged machines are being handled separately in ISO/WD 25785-1, still in development.

So the position of a humanoid in a factory today is this. Its body is abstracted by an openly licensed foundation model. The ERP above it wants to give it work and has eleven partnerships to prove it. Its safety case is half-written and improving. And the interface in between, the one thing that would let a manufacturer buy a robot from anyone and route work to it, does not describe it at all.

This is the part that any excited version of this argument leaves out, so it goes before the recommendation rather than after it. VDA 5050 is published open source on GitHub by two industry associations. Its authors captured no rent from it. MassRobotics is a non-profit working group. ISO 21423 will be a document that costs a few hundred francs to buy and nothing to implement. In the mobile robot category, the interface standard created enormous value and captured none of it.

The money went to the layer above. It went to the orchestrators, InOrbit and its peers, in a multi-fleet orchestration market growing at rates that only look plausible because the base was near zero. It went to the fleet owners who could now buy from more than one vendor. The people who wrote the protocol got a working group and a conference.

So "build a HAL for humanoids" is not, on its own, a plan. It is a public good with a business attached to it, and the attachment has to be designed deliberately. Anyone who tells you the standard is the moat has not looked at what happened to the last one.

Everything above is sourced. What follows is a decision, and it is ours to be wrong about.

One. Write the humanoid profile of an existing interface, not a new interface. The temptation is to design a clean abstraction layer from first principles. That path ends in a specification nobody implements. VDA 5050 and ISO 21423 already carry the hard-won parts, the message topology, the state model, the connection semantics, the factsheet. What they lack is a vocabulary for manipulation work: the task with a tool and a tolerance, the quality result, the qualification a machine holds. That is an extension, it is small, and it is defensible precisely because it does not ask anyone to abandon what they have.

Two. Be in the room while ISO 21423 is still in ballot. This is the timing point and it is the reason this note exists today rather than next quarter. A standard is cheapest to influence before it publishes and nearly impossible to influence after. ISO TC 299 has around 90 participants and a document in ballot expected to land this year. Participation runs through a national standards body, which for us means Standard Norge. Even observer status buys the two things we do not have: the actual draft text, and knowledge of what the next work item will be.

Three. Understand that the product is the capability record, not the protocol. This is the answer to the counter-argument above. The factsheet in VDA 5050 is a machine declaring what it can do. It is trusted because it is simple and because nothing much rides on it. The moment work is routed to a cell on the strength of its declaration, the declaration needs to be backed by something: a quality history, an audit trail, a qualification that somebody will stand behind. A protocol is a public good. A living, quality-fed certification record that makes a capability claim believable is a product, and it is the one this site has been describing since the beginning. Give the protocol away. Sell the reason to believe it.

Four. Build the ERP adapter once, against the standard. The eleven-by-eleven matrix of robot makers and system integrators inside SAP's program is the current state of the art, and it is a bespoke integration per deployment. That is the cost everyone is paying and nobody is naming. An adapter written against a published interface, rather than against a specific robot, is the difference between a consultancy and a product.

Five. Do not build a humanoid, and do not build a foundation model. The floor is being commoditized deliberately by NVIDIA under an open license, and a company with more humanoid engineering than us has just concluded that the body is not the asset. Those two facts point the same direction.

P-12 predicts that a factory OS controlling layout, tools, quality and capacity as versioned configuration becomes central to flexible manufacturing. The evidence here is that the industry agrees and is arguing about who supplies it. P-18 predicts the durable moat is the framework, meaning standards, data, certification and training, rather than the robots. This note is the strongest support that claim has had and also its sharpest warning, because SAP and ISO are both building framework, and neither of them is us. P-10 and P-27 both depend on capability being declarable and trustworthy across an ownership boundary, which is exactly what the factsheet concept starts and stops short of. None of these are scored here, and the register is untouched.

The falsification conditions matter more than the argument, so here they are. If ISO 21423 publishes with a manipulation vocabulary already in it, the gap closes without us and the opening was imaginary. If SAP extends Joule into a published, vendor-neutral interface rather than a partner matrix, the incumbent wins the seam on distribution and there is no room. If InOrbit or a peer ships humanoid support against a public specification within a year, the commercial answer arrives before the standard does. And if humanoids stall in pilots through 2027, the whole question is premature and the interface will be written by whoever is still standing later.

The counter-case to our own recommendation is simply that a company with no factories should not be writing standards for factories. It is a fair objection. The reply is that the parts of this we can test are cheap: a draft profile is writing, and observer participation in a technical committee is a fee and a form. Neither requires a hub. Both stop being available once the document publishes.

SAP News: How SAP brings business AI into the physical world, new robotics pilots across warehouse, manufacturing and inspection use cases ·SAP News: BITZER helps SAP pioneer project embodied AI ·ERP Today: SAP EWM meets humanoid robotics in new Accenture and Vodafone warehouse pilot ·The Robot Report: Humanoid completes live HMND proof of concept with SAP and Martur Fompak ·NEURA Robotics: NEURA, SAP and Bitzer, redefining logistics ·NVIDIA Developer: Isaac GR00T, generalist robot 00 technology ·Sanctuary AI: Sanctuary AI expands physical AI strategy to industrial robotics ·Techcouver: Sanctuary AI evolves model to offer software for other robotics platforms ·VDA: Latest version of the VDA 5050 communication interface for mobile robots has been released ·Fabrico: What is VDA 5050, the standard interface for mixed AGV and AMR fleets ·The Robot Report: VDMA says VDA 5050 version 3.0 will help mobile robot fleets scale ·ISO: ISO 21423, robotics, industrial mobile robots, communications and interoperability ·A3 Automate: ISO 21423, building global consensus for mobile robot interoperability ·InOrbit.AI: InOrbit.AI demonstrates the future of multi-vendor robot orchestration and physical AI at Automate 2026 ·SYNAOS: VDA 5050, MassRobotics, Open-RMF, what is what and what is where ·MassRobotics: Autonomous mobile robot standards published by MassRobotics ·ISO: ISO 10218-1:2025, robotics, safety requirements, part 1, industrial robots ·The Robot Report: ISO 10218 industrial robot safety standard receives major overhaul