Humanoid robot deployment numbers circulate freely and cite almost nothing. A figure appears in a press release, gets repeated in a summary, acquires a citation to the summary, and within a month it is a fact with no traceable origin. Nobody has said anything false, and nothing has been established.
What makes the field genuinely interesting is that there is a concrete, published, technical reason humanoids remain rare on production floors — and it is not the one usually offered. The obstacle is not that the robots cannot walk. They walk well. The obstacle is that the industrial safety standards which govern machines around people assume something a legged robot cannot provide.
The quick answer
The gap between a humanoid demonstration and a production fleet is mostly a functional safety gap. Standards such as ISO 13849-1 and EN 60204-1 are built on the assumption that removing power brings a machine to a safe state. For a walking biped, removing power causes it to fall over — so the safe state is an actively controlled state, which the standards do not contemplate. Until that is resolved, humanoids operate under bespoke risk assessments in restricted cells rather than as certified equipment working alongside people.
Key takeaways
- Ask four questions of any deployment claim: is it autonomous or teleoperated, for how long, how often does a human intervene, and under what safety certification.
- The "fail-passive gap" is the technical crux: a legged robot's safe state requires active control, violating an assumption baked into functional safety standards.
- Recent research validating an industrial humanoid cell did so in a 3m × 1.5m semi-enclosed workspace, and explicitly declined to claim end-to-end certified performance levels.
- Meanwhile 38,000 industrial robots were installed in the United States in 2025, up 11% — overwhelmingly fixed-base arms, not humanoids.
- The published research frontier is dominated by locomotion and manipulation on research platforms, with fall recovery still treated as a feature.
- Mean time between human interventions is the number that would settle most arguments, and essentially nobody publishes it.
Four questions for any deployment claim
Applied consistently, these separate a demonstration from a pilot from a fleet.
| Question | Why it matters | What a weak answer looks like |
|---|---|---|
| Autonomous or teleoperated? | Teleoperation is a person doing the task through a machine | "Supervised autonomy" without a ratio |
| Over what continuous duration? | Ten minutes and ten hours are different engineering problems | Edited video with no stated run length |
| What is the intervention rate? | The single best proxy for production readiness | Not measured, or measured but not shared |
| Certified to what standard? | Determines whether it can work near people at all | "Meets applicable safety requirements" |
The fourth question is the one that most reliably ends the conversation, and the reason is worth understanding rather than merely asserting.
The fail-passive gap
Industrial safety engineering rests on a simple, powerful idea: when something goes wrong, remove the energy. A robot arm that loses power stops. A conveyor that loses power coasts to rest. The machine's safe state is its unpowered state, and every layer of the safety architecture — emergency stops, light curtains, interlocks — is built on that foundation.
A walking biped breaks it. Remove power and it does not stop; it collapses, which is precisely the hazard the system was supposed to prevent.
A feasibility study published in August 2026, Toward Certified Functional Safety for Industrial Humanoid Robots, states the problem directly: "the safe state of a legged robot is an actively-controlled state, which violates the fail-passive assumption" underlying ISO 13849-1 and EN 60204-1. The authors work through the surrounding standards — ISO 13855 on separation distances, IEC 61131-3, the PROFIsafe protocol — and demonstrate an approach on a real cell.
Every emergency stop ever specified assumes that cutting power makes a machine safe. For a humanoid, cutting power is itself the hazard.
Two details of that study are more informative than any vendor claim. The validation was performed on "a Unitree G1 EDU pick-and-place cell in a 3m x 1.5m semi-enclosed workspace" — a small, bounded, guarded space, which is exactly the deployment envelope humanoids currently occupy. And the authors deliberately did not claim end-to-end certified performance levels, instead identifying where the certification boundaries sit.
That restraint is a signal in itself. When researchers working directly on the certification problem decline to claim certification, claims of large certified fleets deserve scrutiny.
What the installation statistics actually show
It helps to put humanoids next to the robots that are genuinely deployed at scale, because the contrast is instructive rather than dismissive.
The International Federation of Robotics reported in June 2026 that US industrial robot installations reached 38,000 units in 2025, an 11% year-on-year increase. Automotive remained the largest adopter at 13,500 units, 1% below the previous year, while the food industry grew 30%. US robot density stood at 307 industrial robots per 10,000 manufacturing employees — eighth globally, ahead of China at 166 but well behind South Korea at 1,220, Germany at 449 and Japan at 446.
Those tens of thousands of machines are overwhelmingly fixed-base arms: bolted down, operating in defined envelopes, certified under exactly the standards the fail-passive assumption underpins. They are boring, and boring is what production automation looks like when it works.
The point is not that humanoids are inferior. It is that the installed base represents a solved regulatory, insurance and integration problem, and humanoids currently do not. A factory adding an arm is doing paperwork it has done before. A factory adding a biped is doing something novel in a domain where novelty is expensive.
Where the research frontier actually sits
Reading recent humanoid literature is clarifying, because it shows what is genuinely hard.
The work is impressive and it is unambiguously about locomotion and manipulation rather than about operations. A September 2026 paper on perceptive general motion tracking reports "zero-shot deployment on a Unitree G1" demonstrating "robust terrain-adaptive locomotion and whole-body motion execution over real-world terrain with obstacles up to 37 cm high, while supporting teleoperation, dynamic motion tracking, and fall recovery".
Read that list of supported capabilities again. Teleoperation is present because teleoperation is still how much useful humanoid work gets done. Fall recovery is present because these machines fall — and a capability list that includes recovering from falls is describing a system whose failure mode is falling over near whatever is nearby.
None of this is a criticism of the research, which is solving real problems well. It is an observation about what stage the field is at. A technology whose current frontier is traversing 37-centimetre obstacles and recovering from falls is not at the stage where fleet deployment claims should be accepted without operational data. The same argument for measuring rather than trusting headline numbers applies here as anywhere — UniverseBlend's case for building your own evaluation rather than accepting published scores transfers to robotics almost without modification.
What a humanoid actually has to beat
The competition is not "no automation". It is a fixed-base arm that costs less, is already certified, and has a supply chain of integrators who have installed thousands of them.
That comparison sets the real bar. An arm bolted to a floor is cheaper to buy, cheaper to insure, cheaper to integrate, and more reliable at a repetitive task than any legged machine will be for some time. Against that, the humanoid's case rests on a single genuine advantage: it fits environments built for people without those environments being rebuilt.
That advantage is real, and it is narrower than it sounds. It pays where the work is varied enough that reconfiguring a fixed cell is expensive, distributed across a space rather than concentrated at a station, and occurring in a facility that cannot be redesigned — an existing warehouse, a retrofit, a site where the layout is constrained. Where the task is repetitive and the layout is changeable, the arm wins on economics that are not close.
So the honest framing of the humanoid bet is not "robots that can do anything". It is that the cost of adapting the world to fixed automation exceeds the cost of building machines that adapt to the world. That is a plausible bet and a genuinely uncertain one, and it turns on integration economics rather than on robot capability.
The compute constraint nobody demonstrates
A second practical issue sits behind the control models. Humanoid control increasingly runs on large vision-language-action models, and those inherit the serving characteristics of any large model.
Control loops need bounded latency. A model producing actions must do so predictably, every cycle, without the tail latencies that are tolerable in a chat interface and dangerous in a machine holding something heavy. That pushes inference on-board, where the power and thermal budget of a battery-powered biped is far more constrained than a rack, or off-board, where a network round trip enters the safety case.
Neither option is unsolvable, and both change the engineering substantially. The relevant hardware constraint is the same one described in why memory bandwidth beats peak FLOPS: on-board inference has no opportunity to batch, so it operates in exactly the regime where accelerators are least efficient. A demonstration with a tethered power supply and a workstation nearby is not evidence about what runs on the robot for an eight-hour shift.
The operational number nobody publishes
One metric would resolve most disagreements about humanoid readiness: mean time between human interventions, measured on a real task in a real facility over a real shift.
It is the number fleet operators track internally for every other class of automation, because it determines staffing, throughput and whether the machine saves labour or relocates it. It is also the number that most cleanly distinguishes the three states people conflate:
- Demonstration — works in a controlled setting, staged, often teleoperated, duration measured in minutes.
- Pilot — runs in a real facility with engineers present, intervention rate high and improving, commercially unproven.
- Production — runs shifts with normal staff, intervention rate low enough that it is not the operational constraint, and the economics close.
Absence of published intervention data is not proof that a deployment is a pilot. Companies withhold operational metrics for ordinary competitive reasons, and that is legitimate. But it does mean that outside claims cannot be verified, and the appropriate response to an unverifiable claim is to treat it as unverified rather than to round it up or down.
If you are evaluating a humanoid for your own operation, ask for intervention rate over a stated period on a task resembling yours, and treat a refusal as information rather than as an obstacle. Retaining that evidence also matters for your own records — UniverseBlend's rundown of the documentation regulators actually ask for is a reasonable frame for what to keep.
The regulatory layer arriving underneath
There is a second certification question forming, and its timing is now fixed.
A humanoid working in a European workplace is a machine governed by product safety legislation, and an AI system operating as a safety component of such a product falls under the EU AI Act's Annex I high-risk category — the tier for products already covered by Union harmonisation legislation requiring third-party conformity assessment. Following the Digital Omnibus amendments, that tier applies from 2 August 2028, as set out in our piece on what the AI Act now requires of deployers.
That date is far enough away to be ignorable and close enough to matter for anything being designed now. It means a humanoid intended for European industrial use will eventually need to satisfy both the machinery safety regime and the AI Act's high-risk requirements — and the fail-passive problem sits underneath both. Solving it is not optional groundwork; it is the precondition for the market existing in its expected form.
What would actually change the picture
Three developments would move humanoids from bounded cells to shared floors, and they are worth watching in preference to demonstration videos.
A standards path for actively-controlled safe states. Either an extension of existing functional safety standards to cover machines whose safe state requires power, or a new standard purpose-built for legged platforms. This is the load-bearing one, and it moves on the timescale of standards bodies rather than of product releases.
Mechanical fail-safety. Designs that fail into a stable posture rather than a collapse — self-locking joints, controlled descent, mechanisms that hold position without active control. This would let legged robots meet the existing assumption rather than requiring the assumption to change, which is by far the faster route.
Published operational data. The first company to publish intervention rates over sustained real deployment will change the conversation permanently, because it will convert an argument about plausibility into a comparison of numbers.
What would not change the picture, despite receiving the most attention: better locomotion, more dexterous manipulation, larger foundation models for control. All are genuine progress on genuine problems. None of them addresses why a certified machine cannot currently be a walking one.
Frequently asked questions
Why can't humanoid robots be safety-certified like industrial arms?
Because functional safety standards such as ISO 13849-1 and EN 60204-1 assume that removing power puts a machine in a safe state. A walking biped that loses power falls over, so its safe state requires active control — a case the standards were not written to cover. This is described in the research literature as the fail-passive gap, and it is an unresolved certification problem rather than an engineering oversight.
Are humanoid robots actually working in factories today?
They are working in bounded, guarded settings — the published feasibility research uses a semi-enclosed workspace of roughly three metres by one and a half. Broader claims may well be accurate, but without published intervention rates, run durations and certification status they cannot be verified from outside, and unverified is the correct way to hold them.
How many industrial robots are actually being deployed?
The International Federation of Robotics reports 38,000 industrial robot installations in the United States in 2025, up 11% year on year, with automotive the largest adopter at 13,500 units. US robot density is 307 robots per 10,000 manufacturing employees, against 1,220 in South Korea. These are overwhelmingly fixed-base arms rather than humanoids.
Is teleoperation a legitimate form of deployment?
It is legitimate and useful, and it is not autonomy. A teleoperated humanoid relocates labour rather than replacing it, which can still be valuable — remote operation of hazardous tasks has clear worth. The problem is only when teleoperated performance is presented as evidence of autonomous capability, since the two have very different economics.
What single question should I ask a humanoid vendor?
What is your mean time between human interventions, on a task like mine, over a full shift, with the safety certification status of that configuration stated? It combines all four questions above and is difficult to answer evasively.
Final takeaway
Humanoid robotics is making real progress on hard problems, and the demonstrations are not fake. What is missing is not capability but the unglamorous apparatus that turns capability into deployable equipment: a certification path, published reliability data, and a safe state that does not depend on the machine staying powered.
The fail-passive gap is the most useful thing to know about this field, because it explains the current deployment envelope better than any claim about the technology's maturity. Watch for movement on standards and on mechanically fail-safe designs. Those will predict production humanoids far more reliably than another video of a robot walking over rough ground.
Sources and further reading
- Ding, Cui, Wang and Wen, Toward Certified Functional Safety for Industrial Humanoid Robots: The Fail-Passive Gap and a Feasibility Study, arXiv 2608.02809
- Li et al., PGMT: Perceptive General Motion Tracking for Humanoid Robots, arXiv 2609.08511
- International Federation of Robotics, US robot industry returns to double digit growth, June 2026
- EU AI Act Article 6, classification rules for high-risk AI systems including Annex I products
