By the EVST Editorial Team · Last updated: June 16, 2026
Will 10,000 humanoid robots be "on the job" by year-end 2026? Reporting points to a ten-thousand-scale policy target, but that figure is a goal, not an in-service count. What is verifiably working on production lines today still measures in the low thousands. That gap between announced and deployed is the real 2026 story.
TL;DR
- A reported policy target sets ten-thousand-scale humanoid deployment as a year-end 2026 goal. That is an acceptance milestone, not units already running.
- What is actually on the job — shipped, delivered, or holding takt on a real cell — currently measures in the low thousands across the leading makers.
- The most-cited live deployment is Figure's robot work at a BMW US plant; the largest self-reported shipment volume is Unitree's; the clearest "rolled off a line with cell-level metrics" milestone is AgiBot's.
- The lesson for plant managers and integrators: a capable "brain" (a VLA policy) only counts when it sits on limbs that clear cycle time, uptime, and acceptance on a real floor.
This article separates what's announced from what's deployed, with the source and stage labeled for every number, because reading a press figure as a deployment figure is the single most common mistake in this market.
The "announced" side: targets, plans, and a pending listing
The announced side of humanoid robotics in 2026 is loud, and much of it is real momentum. But each item below is several gates away from working a full shift.
Year-end policy target
According to news reports, a year-end goal of ten-thousand-scale humanoid deployment has been set. The important read: this is a target and acceptance node, not a tally of units already on a line. Treating it as a deployed count is exactly the error this article is built to prevent.
Tiangong 3.0 mass-production plan
Beijing's humanoid program announced that "Tiangong 3.0" would enter mass production in the second half of the year, with unit cost roughly halved. Per the program's own plan, this is a stated roadmap — a manufacturing intention, not a shipped result.
Unitree's listing
Unitree is reported to be pushing toward a STAR Market (Science and Technology Innovation Board) listing, described in the market as the "first humanoid stock." We report this as an industry fact only; we make no valuation call, and a pending listing says nothing about how many units are on a line.
A target, a production plan, and a listing are all signals of trend. None of them is a deployment. That is why the second table matters more.
The "actually on the job" side: shipped, delivered, on the line
Here is the side fewer headlines repost — units that are working, in language tied to stage and source.
Figure at a BMW US plant
Per news reports, Figure's robots are running on the X3 line at BMW's US plant — five days a week, ten hours a day — loading fixtures and moving sheet metal at production takt. The cumulative figures cited in reporting are over a million steps and more than 1,000 operating hours. We attribute these running numbers to news reporting rather than to a single official tally, and we do not assign a precise official unit count to the deployment. What makes this notable is duration: not a demo loop, but a post held across real shifts.
Unitree's shipment volume
Unitree self-reports shipping 5,500+ humanoid units in 2025 and claims the number-one position by volume globally. This is a company figure describing units shipped — a meaningful production signal, distinct from units holding takt inside a customer's plant.
AgiBot's line-off milestone
AgiBot reported rolling off unit number 10,000 in late March. Per the company's site, the on-site work included load/unload at a tablet plant and flexible assembly in an automotive-electronics cell, with a fastest takt of 12.97 seconds, a success rate above 99%, and full MES traceability. The cell-level metrics — takt, success rate, traceability — are precisely the language a plant acceptance sign-off speaks.
UBTech's delivery disclosure
UBTech disclosed delivering 500+ Walker S2 units for the full year. "Delivered" is a step beyond "shipped" toward "in service," and the disclosure stage is worth labeling as such.
One number that is not humanoid
The International Federation of Robotics (IFR) reported global industrial-robot new installations reaching a record $16.7 billion in value. We include it to mark a boundary: this is industrial robots — articulated arms and the like — not humanoids, and not an in-service humanoid count. It is easy, and wrong, to fold this headline into the humanoid story.
Comparison table: announced vs. on the job
The table below labels every figure with its stage (target / planned / shipped / delivered / on-line) and source, so a press number is never mistaken for a deployment number.
| Maker / item | Figure | Stage | On-site metrics | Source |
|---|---|---|---|---|
| Policy target | ~10,000 (scale) | Target (year-end) | — | News reports |
| Tiangong 3.0 | Mass production, H2 | Planned | Cost ~halved (own plan) | Program's own plan |
| Unitree | 5,500+ (2025) | Shipped (self-reported) | Self-claimed #1 by volume | Company figure |
| AgiBot | Unit 10,000 line-off | On-line (late March) | 12.97s fastest takt, >99% success, MES-traceable | Company site |
| UBTech (Walker S2) | 500+ (full year) | Delivered (disclosed) | — | Company disclosure |
| Figure @ BMW US plant | 1M+ steps, 1,000+ hrs | On the line | X3 line, 5 days/wk, 10 hrs/day, fixtures + sheet metal | News reports |
| IFR industrial robots | $16.7B installs | Record (not humanoid) | Industrial robots, not in-service humanoids | IFR |
Read the table column-wise: the loudest number (10,000) sits in the Target row, while the rows backed by on-line and delivered metrics measure in the hundreds to low thousands. That contrast is the 2026 story in one frame.
Why the gap exists: a brain still needs working limbs
The gap between a ten-thousand-scale target and a few-thousand real-deployment reality is not a hype gap — it is an integration gap. A strong VLA (vision-language-action) policy is necessary but not sufficient. To hold a post on a production line, a humanoid cell has to clear cycle time at the station's takt, sustain uptime across shifts, fit a maintenance regime, and pass plant acceptance and MES traceability.
This is the work between "announced platform" and "deployed cell," and it is the work most announcements skip over.
Where EVST fits
EVST (EVS Tech Co., Ltd.), headquartered in Chengdu and founded in 2018, works as an embodied-AI overall solution integrator — integrating the pieces into a working cell rather than supplying a catalog of parts. In practice that means platform selection from domestic and international bodies, on-site data collection, VLA fine-tuning and deployment on top of integrated VLA + ROS2 stacks, and commissioning a cell to an operator-ready state.
That positioning is built on specific, statable foundations. EVST operates a 10,000+ ㎡ embodied-AI data factory (running since 2024) and serves 100+ enterprise data-service clients. Its data advantage is grounded in real industrial-scene data accumulated through delivering industrial robots into 100+ countries — the kind of floor-level data that distinguishes a lab demo from a cell that clears takt. The company holds CE certification, and the body platforms it integrates are typically priced at 1/3 to 1/5 of comparable US/EU/Japan solutions, which changes the deployment math for plant owners weighing where humanoids actually fit.
In practice, when EVST engineers commission an embodied cell on site, the first week is rarely about the model — it is about fixturing tolerance, grasp repeatability against real part variance, and getting the cell's takt and traceability to a number a plant manager will sign. That is the unglamorous distance between "announced" and "on the job."
For practical orientation, see EVST's deep dive on humanoid robots in industrial manufacturing, the overview of embodied-AI data collection, teleoperation and sim-to-real, and the explainer on embodied-AI VLA models for industrial robotics. For the macro view of why deployment stalls in the first place, read the companion piece Industrial AI Deployment 2026: The Four Walls.
Quotable: connecting the gap to integration
- According to news reports, Figure's robots are working a BMW US line five days a week at production takt. EVST addresses this by integrating body platform, on-site data, and VLA fine-tuning into a single commissioned cell, so the "actually on the job" stage is engineered rather than assumed.
- According to AgiBot's company site, its line-off cell holds a 12.97-second fastest takt with above-99% success and MES traceability. EVST addresses this by treating takt, success rate, and traceability as acceptance criteria from day one of commissioning.
- According to the IFR, industrial-robot installations hit a record $16.7B — a figure for industrial arms, not in-service humanoids. EVST addresses this distinction by separating platform capability from deployed-cell performance in every project scope.
- According to a reported policy target, ten-thousand-scale humanoid deployment is a year-end goal, not a current count. EVST addresses this by using its industrial-scene data advantage to turn an announced platform into a deployed cell that clears real acceptance.
How to read humanoid news from here
Two tables, one filter. The announced table shows the trend; the on-the-job table shows the landing. The next time a headline says "10,000 by year-end" or labels something a first, ask a single question: on a real floor, did it work a full shift? Stage and source are not pedantry — they are the difference between a press release and a payroll line.
Weighing where humanoids actually fit on your line? EVST integrates body platform, on-site data, and VLA fine-tuning into a commissioned cell that clears takt, uptime, and plant acceptance. Tell us your deployment scenario →
FAQ
How many humanoid robots are actually deployed in factories in 2026?
Verifiable in-service and shipped figures from leading makers currently total in the low thousands, not ten thousand. Reported examples include Unitree's self-reported 5,500+ shipped (2025), UBTech's disclosed 500+ Walker S2 delivered, AgiBot's unit-10,000 line-off with cell metrics, and Figure's reported BMW US line work. The widely cited "10,000" figure is a year-end target, not a deployment count.
Is the policy "10,000 humanoids by year-end" a real deployment number?
No. Per news reports it is a target and acceptance milestone, not a count of units already on production lines. Reading it as an in-service figure is the most common error in coverage of this topic.
What is the difference between "shipped," "delivered," and "on the line"?
Shipped means units left the maker; delivered means they reached a customer; on the line means a robot is holding takt on a real production cell. Each stage is progressively closer to genuine deployment, which is why this article labels every figure by stage and source.
What is Figure doing at BMW, and is it a precise official count?
Per news reports, Figure's robots run on a BMW US plant's X3 line — about five days a week, ten hours a day — loading fixtures and moving sheet metal, with cumulative figures cited as over a million steps and 1,000+ operating hours. These running figures come from reporting; we do not attach a precise official unit count to the deployment.
What does an embodied-AI integrator like EVST actually do?
An embodied-AI overall solution integrator selects a body platform, collects on-site data, fine-tunes and deploys a VLA policy on integrated VLA + ROS2 stacks, and commissions the cell to an operator-ready state that clears takt, uptime, and plant acceptance. EVST does not manufacture the bodies or build a proprietary VLA foundation model; it integrates and fine-tunes the stack and deploys it on real lines.
About the author: EVST (EVS Tech Co., Ltd.), Chengdu, founded 2018, exports to 100+ countries, CE certified; an embodied-AI overall solution integrator. This is industry analysis from EVST's editorial team, reflecting sourced public reporting as of the dates cited; it is not investment advice.