Publication date: Friday, 25 September 2026
Author: Malik Carter, Chief Intelligence Architect
Deverout and Associates | Strategic Intelligence for Transformation Leaders
The humanoid robot race is being reported as a capability race. It is actually an evidence race. The organisations that win will not be the ones with the most impressive demo footage. They will be the ones that can prove, with named customers and operating data, that a machine performs a defined task safely, repeatedly, and at a cost the business can justify.
The Signal This Week
September has been the busiest month yet for humanoid robot announcements. OpenAI’s Sam Altman said the company will “definitely do a humanoid,” with no ship date, production target, or manufacturing partner named. Tesla began Optimus production at Fremont in August, but as of September no customer units have shipped, and Elon Musk had already acknowledged in January that no Optimus units were doing useful work inside Tesla’s own factories. XPENG’s IRON moved onto a production line with more than 80% of core processes automated, targeting mass production by year end. Figure AI launched Index, a physical-world dataset that surpassed 69,000 weekly active users within two weeks. China’s TianGong Ultra ran 100 metres in 8.64 seconds. And in July, the US Federal Communications Commission added foreign-produced advanced robotic devices to its Covered List, prompting Unitree to warn that its US sales are at risk.
Individually these are data points. Together they describe an industry where announcement velocity has outrun verification. The strategic risk for enterprise leaders is not that physical AI is immature. It is that the immaturity is being obscured by numbers nobody has audited.
The Claims Are Not the Capability
Search for humanoid deployment figures and confident numbers appear everywhere: Tesla past 50,000 cumulative Optimus units, Figure surpassing 10,000 deployments across partner warehouses, thousands of robots “working” production lines. None of these widely circulated figures originate from the companies they describe. Tesla has never published an audited Optimus production count. On Tesla’s own Q4 2025 earnings call, Musk conceded Optimus was “not in usage in our factories in a material way” and that units were primarily generating training data, not output.
Set against that pattern, Agility Robotics is the outlier worth studying. The company reports its Digit robot has accumulated more than 65,000 operating hours across nine customer facilities, and it names the customers: GXO, Schaeffler, Toyota Motor Manufacturing Canada, and Mercado Libre. Separate reporting puts Agility’s multi-year Digit v5 order book above $300 million, with conversations underway with more than thirty additional prospective customers. The difference is not that Agility’s robots are more advanced. It is that Agility discloses what can be checked — named customers, operating hours, facility count — while much of the rest of the sector discloses what cannot.
| Signal | Widely circulated claim | What is actually verifiable |
|---|---|---|
| Tesla Optimus | “50,000+ units deployed” | Production started August 2026; no audited count; no customer units shipped as of September; Musk conceded in Q4 2025 no material factory usage |
| Figure AI | “10,000+ deployments across partner warehouses” | No company-published deployment count; Index dataset reports 69,000 weekly active users, a data-platform metric, not a robot-deployment metric |
| Agility Robotics | N/A — company reports its own figures | 65,000+ operating hours, 9 named customer facilities, $300M+ multi-year order book |
| XPENG IRON | “Mass production underway” | 80%+ of core production-line processes automated; mass production and store deployment targeted for end of 2026, not yet delivered |
The pattern generalises beyond robotics: adoption claims travel faster than the evidence that would substantiate them, and a leader who builds strategy on the claim rather than the evidence inherits the gap between the two.
The Physical Delegation Boundary
Issue #14 argued that software AI crosses a delegation threshold when it can change the state of the business without a human reviewing every step. Physical AI has its own version of that threshold, and it is stricter, because the actions are irreversible in a different way: a mis-picked pallet can be corrected; a robot arm that makes unplanned contact with a person cannot be undone.
| Level | What the system does | Executive test |
|---|---|---|
| Perceive | Senses and maps its environment | Does it reliably detect people, obstacles, and object state changes? |
| Assist | Performs a task under close human supervision or behind physical barriers | Is the task fenced, caged, or zoned away from people? |
| Coordinate | Shares physical space with people on a defined task | What is the documented safety case for proximity, contact, and recovery? |
| Operate | Performs a chain of physical actions across a shift with limited intervention | Can the organisation name the customer, the facility, and the operating hours behind that claim? |
Agility’s Digit 5 is explicitly designed to work outside fences, sharing space with people rather than behind them — which is precisely why its evidence discipline (named customers, logged hours) matters more than a faster-running robot with no comparable disclosure.
The Physical AI Evidence Stack
| Layer | Diligence question | Minimum evidence |
|---|---|---|
| Deployment status | Pilot, controlled trial, or live production? | Named site, start date, and current operating status |
| Named customer | Who is actually running this, and will they confirm it? | Customer reference willing to be checked, not a vendor case study alone |
| Operating record | How long has it run, and what failed? | Operating hours, incident log, downtime, and recovery rate |
| Safety case | What governs proximity to people? | Documented risk assessment aligned to emerging EU Machinery Regulation and AI Act requirements |
| Regulatory standing | Can this hardware still be sold and supported here next year? | Confirmation the platform is not exposed to export or Covered List restrictions |
| Data provenance | What trained the model behind the behaviour? | Documented source, licensing, and chain of custody for training data |
What the Regulatory Clock Is Doing
The FCC’s July 2026 addition of foreign-produced advanced robotic devices to its Covered List is the first hard signal that hardware sourcing is now a strategic decision, not a procurement one — companies including Unitree have flagged that US sales and future model approvals are at risk. In parallel, the EU’s AI Act and updated Machinery Regulation are expected to require commercial humanoid operators to demonstrate systematic safety cases by Q3 2027, with US OSHA guidance on autonomous robot co-workers expected in H1 2027. Leaders piloting physical AI today are, in effect, choosing a regulatory exposure for 2027 as much as a vendor for 2026.
A 30-Day Physical AI Evidence Audit
- Days 1–5: List every physical AI claim currently informing your strategy — vendor pitches, board updates, competitor benchmarking — and separate the ones with a named, checkable customer from the ones without.
- Days 6–10: For each vendor under live consideration, request the operating record: hours run, facilities, incident and downtime log, not a demo reel.
- Days 11–15: Map where the FCC Covered List, EU AI Act, and Machinery Regulation timelines intersect your intended deployment geography and hardware platform.
- Days 16–22: Commission or review the safety case for any workflow where the system will share physical space with employees or customers.
- Days 23–26: Trace the training data behind the vendor’s model — provenance, licensing, and whether it was collected on the same hardware platform you intend to deploy.
- Days 27–30: Decide: proceed on verified evidence, pilot narrowly to generate your own evidence, or wait for the regulatory and disclosure picture to clarify.
The Bottom Line
The question for transformation leaders is no longer, “Is physical AI ready?” It is: Whose numbers can actually be checked, and what does our own evidence say once we stop borrowing theirs?
The organisations that will out-execute this cycle are not the ones with the most dramatic robot footage. They are the ones building the same operating discipline in hardware that Issues #11 through #14 argued for in software — named accountability, logged evidence, and a tested intervention path, before the marketing claim is mistaken for the deployment.
This weekly publication continues the argument in real time. Return to deuerout.com for the next issue, updated signals, practical frameworks, and the questions transformation leaders should be asking before capability becomes dependency.
Sources and attribution
- Objectways, Humanoid Robots News Roundup for September 2026, 23 September 2026: source.
- Technology.org, Humanoid Robots in 2026: What Is Actually Deployed, 18 July 2026: source.
- Gartner, Gartner Predicts Fewer Than 20 Companies Will Scale Humanoid Robots for Manufacturing and Supply Chain to Production Stage by 2028, 21 January 2026: source.
- Robotics Center of Silicon Valley, State of Robotics 2026 Report, regulatory timeline (EU AI Act / Machinery Regulation, US OSHA guidance): source.
Editorial note: Company-reported performance figures, vendor case studies, and unverified secondary-source deployment counts are identified as such throughout and should not be treated as independent benchmarks.
