The robot is not the product: the factory of the future

Zhaopeng Chen read Isaac Asimov as a kid, and he never really stopped. Maybe that's why Agile ONE, the humanoid his company started building this year in Bavaria, looks vaguely like the machines in the Will Smith adaptation of I, Robot.

"Every single person on the planet will get a personal robot," the founder and CEO of Agile Robots told me in Munich last week. "This is not going to be very far away. It's not in days, but in years, decades, it will be there."

Eight billion robots, give or take.

Wow.

What makes it interesting, and maybe even credible, is the roadmap underneath it. Because Chen's company is doing almost nothing that looks like building a consumer robot. Agile Robots sells force-sensitive arms, dexterous hands, autonomous mobile transport, automation cells, and a software layer that stitches them together. It has more than 20,000 automation systems deployed. It has bought four systems integrators in three years. Its humanoid is aimed at machine tending and material handling in factories, not at your kitchen.

(Yet.)

That gap between the science-fiction endpoint and the industrial roadmap is the most useful thing to study about this company. Chen argues it isn't a detour. It is, in fact, the only viable route.

Two viral loops: technical and business

Ask most humanoid CEOs why they'll win and you get data, models, or capital. Chen's answer is structural, and I haven't heard it framed this way by CEOs in the U.S. or China.

Large language models improve, he says, because the growth engine is understood, mapped, and fundamentally simple — it's entirely digital and it sits on internet-scale text.

"You just create the whole channel — model, data, compute and outcome," he told me. "All you need to do is basically pump in money. So if you're pumping money, then the real results will come out. And this flywheel is working really fine."

Physical AI isn't that.

"You really have to create the model, the body, physical part, sensory system, whatever, and then application. And only when application and the valuable applications are massively deployed, then this will come back, drive capital, drive the compute, and drive the body and drive the application again."

Which leads to the thesis:

"For me, the wheel is two levels, two layers, let's say. Layer one is technical. Layer two is also business. If you cannot close that, it doesn't matter how much capital you have."

Two loops, not one. The technical loop is the one everybody talks about: deployments generate data, data improves models, better models enable more deployments. Agile Robots' own marketing calls it the Physical AI Flywheel, and every serious player in this space has a version of the same diagram.

The business loop is the one that gets skipped. Deployments have to generate margin, margin has to fund the next generation of hardware, and the hardware has to be cheap enough that the deployment made economic sense in the first place. A technical flywheel spinning inside a business that never closes is just a very expensive research program.

That's a pointed thing to say in September 2026. Humanoid companies raised $2.7 billion in August alone, most of it in China. Figure has committed $3.5 billion for AI compute against roughly $1.9 billion raised. Neura Robotics, an hour up the autobahn in Metzingen, raised up to $1.4 billion in June.

Chen says Agile Robots has raised roughly $1.5 billion over seven years, is backed by SoftBank, and isn't losing sleep.

"Competition is always good. I'm trained in a way to compete," he said. He'll concede the American advantage — "US companies, they have certain … it's very unique, unfair sort of advantage. That's something I have to admit." He won't concede that money alone decides it. "I don't think any business will be sustainable by only pumping in money."

Nobody's factory will have 10,000 humanoids

Here is where Agile Robots diverges hardest from the pure-play humanoid startups: the humanoid is not the product. No individual robot is the product.

The factory is the product.

"The goal is to solve the problem, and the second thing is to solve it with the least amount of automation as possible to give them the best ROI," Rory Sexton, Agile Robots' executive director, told me during a day at the company's Munich facilities. "Sometimes you just need a smart robot, not a very complex robot."

Sexton spent 20 years at Apple. He joined the day the iPod launched, and by the time he left he was responsible for 120 plants and 2.5 million workers. He came out of retirement for this. His read on where humanoids sit is refreshingly unromantic.

"Humanoids, I think for many years are not going to operate independently," he said. "They'll be part of a set of solutions that go in."

Think about what that implies for product strategy. If the humanoid is one component in a heterogeneous cell — conveyors, fixed arms, AMRs, vision, a human or two — then the thing you're actually selling is the integration. The humanoid is a line item.

So Agile Robots owns the whole stack:

• Arms. The FR3 force-sensitive research platform (via Franka Robotics), the Diana 7, the Thor series.
• Hands. The Agile Hand, which is the crown jewel and gets its own section below.
• Mobility. idealworks AMRs and AGVs, already deployed and integrated in customer plants.
• Humanoid. Agile ONE.
• Software. AgileCore, which handles perception, planning, control, and fleet and system orchestration across all of it.
• Integration. BÄR Automation, XNG in India, and thyssenkrupp's former automation engineering arm, now rebranded Krause Automation — roughly 650 people and ten sites across Europe and North America.

"You've got to look at the problem and then go: what are the parts of physical AI and normal automation?" Sexton said. "You've got to merge those together to bring one solution."

The integration piece matters more than it looks. Agile Robots' mobile platforms are already wired into customers' manufacturing execution systems — the software that tells a plant what to build, when, and where the material is. That's a multi-year enterprise sales cycle that most humanoid startups have not begun.

"Putting a semi-humanoid on top of it keeps it integrated," Sexton said. "All the route planning, the task planning, it's just another part of the system."

A humanoid that arrives as a new node in an already-orchestrated fleet is a procurement decision. A humanoid that arrives as a novel category requiring its own control stack, its own safety case, and its own MES integration is a project.

Who else actually thinks this way?

It's tempting to frame this as Agile Robots against the field. That's not really true.

Neura Robotics is the closest structural analogue, and it's German too. Neura sells the 4NE1 humanoid and 4NE1 Mini, the MAiRA cognitive cobot, the LARA arm, the MAV+ mobile transport, and a quadruped, then connects them through Neuraverse, a cloud platform where deployments feed a shared pool of physical intelligence, trained partly in a network of physical facilities it calls NEURA Gyms. Same multi-form-factor logic, same data-flywheel argument.

Agility Robotics has Agility Arc, a cloud platform for orchestrating and monitoring Digit fleets. That's a real acknowledgment that the robot alone doesn't ship, but it's orchestration for Agility's own robot, not a general automation business.

UBTech sells Walker S2 into Chinese automotive plants with a coordination layer across multiple robots, and has the manufacturing base to back it.

Figure and 1X are counter-examples, deliberately. Figure's bet is Helix plus BotQ: build the general-purpose robot and the manufacturing capacity to make a lot of them, and let the generality collapse the integration problem. 1X is aimed at the home, which is a different business entirely.

So the distinction isn't "systems thinking versus robot thinking." 

Several companies are doing systems thinking. The distinction is how you get the systems business.

Neura and Agility built platform layers. Agile Robots bought integration companies with existing revenue, existing plant relationships, existing certified installations, and existing engineers who know how to quote a project. Krause alone brings a business founded in 1950 that has been putting automation into automotive plants for decades.

That's the actual bet: that in physical AI, distribution and integration are harder to build than models, and cheaper to buy now than later. Whether it works depends on whether the acquired integrators can sell the new stuff or merely keep selling the old stuff. Roll-ups in industrial automation have a mixed history for exactly that reason.

It also has a cost the pure-plays don't carry. 

Agile Robots did roughly €200 million in 2024 revenue at about 32% gross margin — respectable for a systems integrator, thin compared to what a software-defined robot company hopes to earn. Sacra pegs 2025 revenue at $339 million with 2026 guidance around $690 million, and profitability targeted for 2028–2029. 

Chen's business loop closes, but it closes on integrator economics.

The hands

If you want to understand why Agile Robots believes it can compete on hardware, look at the hand. Hands are the hardest part of a robot. It's the part of the humanoid stack where almost everyone gets stuck, and it's the part Chen personally built his career on.

The Agile Hand runs 21 joints and 21 degrees of freedom in a 1.5 kilogram package, with five identical modular fingers matched to human hand proportions in joints, size and range of motion. Joint torque and position sensors sit in every actuated joint. 

Fingertips carry tactile sensing. Joint speeds hit 360°/s.

Unpack those numbers. Twenty-one DOF per hand is near the top of what anyone ships: most industrial humanoids run six to twelve, and plenty run a two-finger gripper and call it a hand. The 1.5 kg mass matters more than the DOF count, because every gram in the hand is mass at the end of a moment arm, and it taxes the arm, the shoulder, the balance controller and the battery. Torque sensing in every joint — not current sensing inferred from the motors — is what gives you real force control, which is what lets a robot insert a connector without snapping it, or pick up a raw egg without smashing it.

The modular-finger design is a manufacturing decision as much as a mechanical one. Five identical fingers means one part number, one assembly process, one spare, and a yield curve that improves five times faster per hand built.

It's tempting -- and I've seen it from Genesis AI and others -- to have individual finger lengths. But it does come with a maintainability cost.

This traces directly to DLR. The German Aerospace Center's Institute of Robotics and Mechatronics in Oberpfaffenhofen is one of the most consequential robotics labs in the world — it produced the torque-sensing lightweight arm lineage that became KUKA's LBR iiwa, the Justin humanoid, and the DLR Hand series that defined the field. Chen led hand development there before spinning out in 2018. Franka Robotics, now an Agile Robots subsidiary, comes out of the same lineage.

The company holds 650+ patents. Its own factsheet claims a top-ten global position for humanoid patent strength; LexisNexis Patent Asset Index ranks it as #7 among dedicated humanoid startups.

Agile ONE, by the numbers

174 cm. 69 kg. 20 kg payload. 2.0 m/s walking speed. Bipedal. Two Agile Hands, 21 DOF each. Intelligent cameras, LiDAR, 360-degree multi-modal sensing, proximity sensors, speech recognition, a rear-view camera, expressive eye rings, and a chest information display for status and intent. Built in-house at Fürstenfeldbruck, near Munich.

Third-party trackers put total DOF around 40, battery runtime near three hours, compute on NVIDIA Jetson, and price around $100,000. Agile Robots confirms none of that, and says it doesn't release pricing: treat those figures as unconfirmed.

The control architecture is layered into three tiers: strategic planning and reasoning, rapid dynamic response, and low-level fine manipulation. Slow thinking, fast thinking, and reflex. That's converging industry-wide — it's structurally what Figure's Helix does with System 1 / System 2 — but the fine-manipulation tier is where 21-DOF torque-sensing hands actually pay off, because the reflex layer has something worth sensing.

On the model side, Agile Robots signed a partnership with Google DeepMind in March to bring Gemini Robotics foundation models onto its hardware, joining Boston Dynamics and Apptronik on DeepMind's roster. The company's own published evaluations, using its FR3 Duo dual-arm platform, report 89.6% success on precision part insertion, 78.9% on diverse tool kitting, and 74% on general pick-and-place. Those are company-reported figures, not independently verified, mind you.

Data comes from three streams: real-world data from the installed base and Agile Robots' own manufacturing, structured teleoperation data from its Robots Academy in Munich (head- and wrist-camera capture of human task performance), and large-scale simulation and visual data.

Compute is a sovereignty play. Agile Robots is an anchor customer for the Deutsche Telekom–NVIDIA Industrial AI Cloud in Munich, Europe's first industrial AI cloud. For European industrial customers nervous about where their production-floor data lives, that's a real differentiator, and one no American humanoid company can offer them.

What the factory floor actually says

I toured the production lines. Some of the work of building robots is being done by robots.

At least two lines — arms and torsos — are fully automated, as is testing. Hands and final assembly are done by human technicians, which is a nice irony: the machine designed to replace human hands still needs them to get made. That's a reasonable read of the difficulty curve. Arms and torsos are large, rigid, well-toleranced assemblies. A 21-DOF hand with sensors in every joint is a watchmaking problem.

Sexton says the plan goes further: "We don't need people to build arms or torsos. We will have some dark factories."

The throughput numbers, from the walkaround:

• Thousands of robot systems shipping from German operations this year
• Believed capacity of 8,000–9,000 per year, with full capacity targeted next year
• Current throughput around 10 units a day, at three to four days per robot
• Roughly a two-month hardware upgrade cycle
• Some customer orders arriving in increments of hundreds

The gap between 10 a day and 8,000–9,000 a year is not trivial — call it 2,500 units at current run rate against a target roughly three times that. 

Supply chain is deliberately multi-sourced: multiple battery suppliers, multiple motor-actuator suppliers, no single-vendor dependency and no attempt to vertically integrate everything. Manufacturing will be regional — Asia and America as well as Germany — which reads as tariff and sovereignty hedging as much as logistics. 

But the German component supply chain stays central. "A lot of people want to get into this supply chain," Sexton said.

The unit economics

Agile Robots doesn't publish pricing. Sexton offered the yardstick customers use instead: two shifts of labor in China run about $80,000 a year, and industrial buyers want a two-year payback.

That brackets the viable price for an industrial humanoid at roughly $160,000 — if the robot can actually replace two shifts, at human speed, with acceptable reliability. 

Which is why the delivery model may end up mattering as much as the hardware.

"You could do a leasing model," Sexton said. "I've got Christmas coming. I want another 50 of them, please, and you take them back come December."

Robots-as-a-service only works if the fleet is genuinely general. A unit leased for Q4 holiday fulfillment has to earn its keep in Q1 doing something else — cabling a data center, kitting tools, tending a machine.

"You need flexible robots that can do multiple things," Sexton said. "The more flexible the robot, the more useful it is. And hence the humanoid discussion."

That's the cleanest justification for the humanoid form factor I've heard from an industrial vendor. Not that legs and arms are optimal for any given task — they aren't — but that a fleet you can redeploy has a utilization curve a fixed cell can never match. Sexton's endpoint: "Manufacturing as a service, or service as a service, is coming."

Reconfigurability is the same argument at plant scale. Today, changing a production line is a months-long project. "There's going to be a portion of the factory where you can redeploy what your machines, your robots are doing," Sexton said. "If you come in with a new model, it's not such a big deal."

And the demand pressure is concrete, not theoretical. Agile Robots has a Franka plant an hour from Munich that's selling, in Sexton's words, like hotcakes. Adding capacity means hiring 40 to 50 trained workers, which will take many months. "You don't need higher-order math to see that the demand is there."

The regulatory wall

Agile ONE walks. That is, right now, a problem.

"There is no regulatory environment in the West, Europe and North America, for legged, bipedal humanoids working with humans," Sexton said. "There's no standards for this yet."

Agile Robots' wheeled platforms are already certified for human-collaborative work. The bipeds are not, because the standard doesn't exist: there is no ISO 10218 or ISO/TS 15066 equivalent covering a 69 kg machine that can fall over. The company says it's working with the agencies to help write those standards.

This is the most underrated gating factor in the entire Western humanoid sector, and it's a genuine structural advantage for a company with a certified wheeled fleet already in plants. Agile Robots can deploy a semi-humanoid upper body on an idealworks AMR base today, generating data and revenue on the technical and business loops both, while the bipedal certification question resolves on someone else's timeline. Competitors whose only product has legs are waiting.

The United Nations advantage

Both executives kept circling the same idea from different directions: that a structural advantage of this company is that it isn't from anywhere in particular.

"In terms of the company makeup, we're sort of like the United Nations," Sexton said. German, Indian, European and Chinese in roughly equal measure, more than 60 nationalities across 3,500-odd people, over 2,000 of them in R&D.

"We don't have a German culture, per se. We have an Agile culture, and that's very important. The Chinese are unbelievably fast at supply chain and the art of the impossible. Germany is very, very good at design and deep mechatronic knowledge. On the AI side, Germany is very good, but clearly California and America, that's very strong. If you're going to build a successful company in this, you've got to take it all."

Chen frames it as Europe's second chance. 

He's blunt that the continent lost round one — Europe and Germany, he said, got lost in the last foundation model competition. He thinks this round is different. "Europe stands a chance for the physical AI, because it is not only software. It is heavily, heavily software and hardware in combination."

He also thinks the walls going up around this technology are predictable and temporary. 

Agile Robots didn't go multinational to hedge geopolitics — it was built that way from day one — but he's seen this movie. "If you look at the history of the last 100 years, this is happening and happening again and again and again for every new technology." China once made Volkswagen take a joint venture partner. Everyone protects, then everyone exports. "This is never-ending. It's just history."

The bet

Chen's other literary reference is Jules Verne, who described the submarine roughly a century before anyone built one. There's a line from Verne posted at the Paris airport that Chen quotes as his operating principle: whatever mind can imagine, mind will realize it.

Eight billion personal robots is a lot to imagine, and the obstacles between here and there are as much economic as technical ... a robot for every person on the planet runs into cost curves and income distributions long before it runs into physics.

But the argument Agile Robots is making deserves to be taken seriously on its own terms, separately from the prediction. It says the humanoid is a component, not a category. It says the buyer is purchasing an outcome, not a robot. It says the winning position is the one that owns the integration, because that's where the customer relationship lives and where the deployment data comes from.

If that's right, the scoreboard everyone is watching — humanoid units shipped, funding raised, demo videos posted — is measuring the wrong thing. The number that matters is how many production lines are running, how much margin they throw off, and whether that margin funds the next generation of hardware.

Chen's version: two loops, both closed. Right now the company is closing them by peeling protective film off smartphones, one line at a time. And other jobs, of course -- like building more robots -- and increasingly adding up to a full-feature smart factory of the future.