Humanoid hype, read from the outside: what a physical-AI primer adds to the PLAZA question

The announcement at ITB Berlin in early March 2026 was a partnership between the Heilbronn hotel group PLAZA and the Metzingen manufacturer NEURA Robotics to bring the 4NE1 humanoid into reception, guest service and housekeeping, positioned by both sides as support for staff rather than replacement. The August article on this site did what could be done with that material: it separated the announced pilot from a verified rollout, it put the partners' own labour-shortage estimate next to the Bundesagentur's registered vacancies, and it compared the payback arithmetic of a humanoid at a reception desk with that of a wheeled robot in a service corridor. It closed on a list of what an operator would need to see before drawing conclusions. As of late August none of that had been published, and this article has found nothing new since.

What that piece could not do was step outside the announcement. Everything it had on the humanoid category came either from the two partners or from this site's own experience with wheeled machines. That is a legitimate basis for scepticism but a narrow one, and a reader was entitled to ask whether the scepticism was an integrator's reflex. An outside source that is not selling a humanoid, not buying one and not competing with one is the thing that was missing.

The report mentions no German manufacturer, no hotel and no hotel pilot. Its examples are Tesla's Optimus, 1X's NEO, Figure's funding round and a handful of research groups. That distance is the point. Nothing it says about humanoids was written with the PLAZA pilot in mind, so nothing it says can be read as commentary on it. It can only be read as a description of the category into which that pilot falls.

Frank does not hide his sources' scepticism, and he places it early. He writes that several of the experts he spoke to for the report indicated that the hype building around physical AI is louder than what is currently justified. A page later he describes the view that machines are on the cusp of taking over an almost limitless addressable market for labour, and observes that this narrative is especially loud in humanoid robotics, where the standard argument runs that a machine shaped like a person can step straight into buildings built for people. The expected reward for that, he notes, is what is currently fuelling the sector's valuations, and he cites Figure's billion-dollar financing round as the example.

Two features of that passage matter for an operator. It is an account of interviews, and Frank does not name the experts or count them, so it cannot be weighed as a survey. And it is a statement about the narrative, not about any machine: the claim is that the story of humanoids replacing labour runs ahead of what has been shown, not that humanoids do not work. Read with that precision, it lines up with what the August piece observed from the inside, that the announced hotel roles are a bundle of reception, service and housekeeping tasks, and that no evidence has been published of a humanoid carrying that bundle in a live property. The primer says the same about the category from outside it.

The hardware chapter of the report is the part that a hotel operator is least likely to have heard and most likely to find useful. Frank quotes the robotics investor Jack Pearson, who says that if you ask a humanoid company what is holding it back you hear AI, data and software, and if you ask its engineers off the record you hear heat: actuators overheat and the machine folds. Pearson's phrasing, which the report sets out as a pull quote, is: ‘The dirty secret of humanoid robotics isn't intelligence. It's thermodynamics.’ Frank's gloss is short. Actuators generate friction, friction generates heat, many of today's research demonstrations run for minutes, and a production robot has to run for hours without overheating.

He then makes the arithmetic visible. As a humanoid gains range of motion, its actuator count rises: he gives Tesla's latest Optimus as 28 actuators driving 50 degrees of freedom. He notes that 1X's NEO takes a different route, moving actuators to the centre of the body and transmitting force through cables to save weight, and adds that nobody yet knows whether that architecture holds up in real environments at scale. The report treats these as design bets being fought out in internal engineering meetings, not as settled products.

For an operator this translates into one question, and it is not the one usually asked. The question is not what the machine can do in a demonstration but for how many consecutive hours it can do it under load in a warm lobby in July, and what it does when an actuator approaches its thermal limit: slows, stops, or hands over. A pilot that reports task success without reporting continuous runtime has reported the demonstration, not the shift. The August article asked for unassisted uptime as one item among six. The primer explains why that item is the one most likely to separate a humanoid from a wheeled machine, whose drivetrain does not have the same problem.

The report's chapter on the deployment gap contains a sentence about pilots that the August piece could have used and did not have. Most robots in production today, Frank writes, are programmed for narrow, repeatable tasks and remain the classical machines of the previous paradigm; the early pilot programmes now under way at some firms mostly involve simple repeatable tasks as well, which, in his words, invalidates much of the benefit of general-purpose intelligence. He is not criticising the pilots. He is observing that an organisation absorbs a new class of machine at its narrow end, and that the narrow end is where a general-purpose machine has least to offer over a specialised one.

Applied to a hotel, and only as a question, that reads as follows. If the first tasks a humanoid takes on in a property are the repeatable ones, carrying items along a corridor, delivering to a room, moving linen, then it is doing wheeled-robot work at humanoid cost and humanoid thermal budget, and the pilot will tell the operator little about the general-purpose case it was announced for. If it takes on the unstructured tasks, the reception bundle, then it is testing the thing the primer's experts consider overhyped, and the results will be worth having precisely because they will be rare. Either outcome is informative. What an operator should not accept is a pilot report that blurs the two, counting corridor runs as evidence for the desk.

Two shorter observations in the primer are worth a paragraph each for anyone hosting a pilot. The first is about remote hands. The report describes teleoperation, a remote person taking over when the machine meets something it cannot handle, as likely to remain a standing part of deployed physical AI, and gives as its consumer example 1X's NEO, a home humanoid that is being sold on the basis that it will be, at least sometimes, operated by remote staff. Whether any given hotel pilot works that way is unknown and this article does not assert it. The question it raises is simple: if a remote operator can take over a machine standing in a hotel lobby, the operator sees the lobby, and the hotel should know who that is, where they sit, and what the recording rules are, before the first guest walks past.

The second is about why manufacturers pilot at all. The report argues that the real contest in robotics is the race to gather enough real-world data to train the next models, and that this creates a strong incentive to ship hardware into live environments as early as possible, because deployments generate the data that improves the model that wins the next deployment. Frank presents this as the central commercial logic of the field, not as a criticism of anyone. The August article noted NEURA's own statement that its systems are prepared for hotel tasks in its training centres with hospitality practitioners. The primer's point sits beside that: a live pilot is, for any manufacturer in this field, also a data source. A host property is entitled to ask what it receives in return, and to settle in writing what leaves the building.

What the primer adds to the August reading is a witness. The scepticism about humanoids at a reception desk no longer rests only on an integrator's experience with wheeled machines; it now has beside it a report written by someone enthusiastic about the field, citing interviews with people inside it, saying that the narrative runs ahead of the evidence and that the binding constraint is thermal before it is cognitive. That does not make the PLAZA pilot a bad idea. A pilot is how the evidence gets made. It makes the pilot's eventual report the thing to wait for, and it says what that report needs to contain to count: continuous hours under load, not task demonstrations; the split between repeatable and unstructured tasks; who holds the remote link; and what data leaves the property.

What does not change is werob's position, and it is worth stating plainly so that the primer is not read as saying more than it does. werob is a manufacturer-independent systems integrator. For transport, delivery and cleaning tasks in hotels and care facilities it specifies wheeled, task-specific machines, because those are the ones with calculable payback and no thermal budget to manage. It will specify a humanoid when a live site, not a demonstration, shows one carrying an unstructured task for a shift. The report supports the reasoning behind that position. It says nothing about werob, measures nothing werob does, and this article does not borrow from it any claim about werob's own methods or results.

Everything attributed to the primer comes from the 25-page PDF of A Business Leader's Introduction to Physical AI by Aaron Frank, dated March 2026, read in September 2026, and is paraphrased except for the two-sentence Pearson pull quote and the phrase ‘louder than what is currently justified’, both reproduced as the report prints them. The expert scepticism is Frank's account of unnamed interviews and cannot be weighed as a survey. The actuator and degrees-of-freedom figures for Optimus are the report's and have not been checked against Tesla's own material. The remarks on 1X's NEO and on teleoperation are the report's descriptions of that product, not of any machine in a German hotel. The status of the PLAZA and NEURA pilot is as this site last recorded it in late August 2026: announced at ITB, no named properties, dates or operating figures published; a web search on 16 September 2026 surfaced only the March coverage and no operating results, which is not the same as there being none. Nothing on this page is a measurement of the 4NE1, of NEURA's training facilities, or of any humanoid in operation, and nothing on it describes werob's own tools or results.