Putting one humanoid robot to work in a factory is a pilot. Exciting, maybe, but not life-altering. Putting 1,000 to work is a completely different animal, and that’s the challenge Hexagon Robotics signed up for in April when German motion technology company Schaeffler committed to deploying at least 1,000 of its Aeon humanoids across its global factory network by 2032.

But according to Hexagon Robotics president Arnaud Robert, the math is not what you’d think.

“It’s not one robot doing one task that you multiply by a thousand,” Robert told me on my Humanoid Daily podcast . “That would be linear, but not that exciting.”

It would also be easy … one plus one plus one. Instead, the starting point is thinking of the robots as a fleet and not individuals, Robert told me.

“I think it’s really the fact that now you have to think in terms of a fleet of robots and how you would deploy a fleet of robots across multiple factories, not just one, in an optimal way,” he told me. “Factory floors are quite complicated. There’s demand in a particular area of the factory; it can move to another area of the factory, and so on.”

That makes sense, but it’s important for a reason that’s non-obvious: it explains why specifically humanoid robots exist in the first place. If a robot does a job in a specific spot, you don’t really need a humanoid. Instead, you just need a conventional automation cell, maybe a robotic arm. A humanoid is for complex multi-step jobs and processes that involve dextrous hands but also free and sometimes non-deterministic movement. Hexagon designed Aeon, which it launched in June 2025 , to handle manipulation like pick-and-place and moving boxes, but also part inspection and reality capture, which is Hexagon’s core business as a measurement technology company.

The payoff, Robert says, comes when something breaks down on the line.

“A bottleneck could be because you have only three people that can do inspection … One of them is sick, one of them is on vacation, and then you have only one left. So what do you do?” His answer, just as if the humanoid robots would be humans, is to redistribute labor on the fly: “If there’s a bottleneck in inspection, you can send 50 Aeons to get rid of that bottleneck and the factory line is back running.”

Even a single task can theoretically justify a humanoid if it requires both mobility and dexterity. Hexagon is piloting Aeon at BMW ’s Leipzig plant, and one job there is moving a spoiler from one line to another. That’s dexterity – picking up the spoiler in a way that doesn’t damage it – plus mobility. And flexibility, in case the pick-up or drop-off points change slightly.

This is a very different vision than conventional deterministic automation. And it uses the capability of a humanoid – whether wheeled or legged – to plug gaps wherever they exist.

Interestingly, Aeon is sort of a hybrid humanoid robot: it has two legs, but each one ends in a wheel instead of a foot. While sometimes seen in quadruped “dog” robots, this is incredibly unusual in humanoids. Manufacturers like Humanoid AI, Agibot and Genesis AI have wheeled-based humanoids without legs, but as far as I know, only Agility Robotics’ just-announced Digit 5 may come in a wheeled bipedal design as well (a wheeled version appears in the company’s launch video).

Hexagon tested bipedal designs and several wheel designs before settling on this one, and the reason is mostly about how factory work actually happens, the company says. Unsurprisingly, most jobs involve getting from one place to another, and on flat factory floors wheels are faster and use less energy than walking. The legs are what keep the wheels from being a limitation. The wheels give Aeon the benefits of speed and efficiency, while the legs allow the robot to climb over objects and even go up stairs, which is the usual argument against wheels on a robot.

“It turns out, actually, that even going upstairs, for example, is very efficient with wheels, because you can use the inertia of the wheel,” Robert says. “As it goes into one stair, it starts rolling and it kind of moves the body forward, and then so on.” So Aeon gets wheel-speed mobility across the factory and leg-style access to the parts of the factory that aren’t flat.

BMW says Aeon can roll at up to 2.5 meters per second.

Interestingly, Robert told me that legs actually win when the robot is standing still to do a task: “in terms of power management, bipedal is actually way more robust if you’re standing still to do a task.” But a robot that only stands still probably doesn’t need legs at all, and Hexagon built its own power management system to make the wheels work.

Hexagon is doing something interesting with training data for Aeon. Of course, every humanoid robot company in the world is scrambling for physical AI training data from a multiplicity of sources: simulation, synthetic video, teleoperation, sensor gloves, head-mounted cameras on people. Hexagon uses all of those, Robert says, but two other things are necessary.

The first is customer-specific data. A task at BMW or Schaeffler is very specific to that factory, and that data is not something either company wants floating around in public. So Hexagon had to figure out how to train on customer data while protecting it.

The second is what Robert calls sim-to-real-to-sim.

“It’s not really sim-to-real," he said, speaking about training data. "It’s sim-to-real-to-sim. And it’s that last loop that is really interesting.”

Simulation data is critical for robotics. It’s manufactured in digital worlds that recreate, as closely as possible, the real world environments robots will be working in. But there’s always a sim-to-real gap.

Thanks to the fact that Aeon is loaded with sensors (22 of them, Robert has said previously ), it records not just the task it is working on but the entire environment around it. That matters because factory floors are messy: “A human could be walking by, a tray is in the way, the part was supposed to be on the left, now it’s on the right because somebody placed it there.”

All of that goes back into training. “So as we actually do the task itself, we feed it back into the training,” Robert says.

Which is kind of a virtuous circle for training data: the robot doing the work is also the robot building the training dataset. Even better, the training data isn’t just video. You have the entire robot’s entire sensorium attached: grip force, hand position, the small corrections it made along the way. That’s a far richer source than hours of unlabeled video. And Robert thinks volume is becoming less important anyway, as training methods improve: “you actually not only need less data, you may need different types of data.”

But it will still be a challenge to scale from pilots to 1,000 robots in a company’s factories and warehouses.

“A pilot demonstrates feasibility, but when you go to production, you need two things: repeatable performance and scale,” Robert says. “You can do it once, you can do it twice. But when you go to production, you have to be able to do it a hundred times in a six-hour shift with a high level of performance.”

To improve both repeatable performance and scale, Hexagon boils all its metrics down to two. The first is cycle time, defined by the customer’s shift: “it’s really about: we have a shift of six hours and, as an example, 500 parts need to be moved in that six-hour shift. That’s our metric.” That’s a different metric than raw speed per action, mostly because today’s humanoids are still speed-challenged.

“The speed of movement is just slower than a human worker who has been doing it for the last five years and knows exactly the positioning and so on,” Robert admits. "That will change over time, but currently that’s the case.”

The main point of the metric, however: if a robot hits the shift target, the occasional miss or slower motion doesn’t kill the business case.

The second metric is human intervention, and it’s tied to a strong position Hexagon has taken on autonomy.

“We don’t believe in a factory you can do teleoperation,” Robert says. “You cannot be autonomous and then also have frequent human intervention.”

This is where it helps that Hexagon’s parent company has decades of industrial experience. I’ve seen this pattern with Apptronik and Agile Robots too: companies that come from industry and have a long background in automation – not just humanoid robots – tend to understand what production needs better than first-principles startups building robots from scratch because the founder thought it was cool.

One factor that will help is Schaeffler’s Humanoid Gym in Germany, which Robert describes as “an exact replica of the production environment, but where you can train the robot.” Aeon entered the gym earlier this year.

The gym serves two purposes. Train a robot well there and it should work as-is on the real line. And if you nail it, Robert says, “you can not only train one Aeon, you can, say, train a hundred Aeons, a thousand Aeons, through basically replicating what you’ve learned on one to all the other ones.”

Hexagon is testing multiple use cases in the gym, and not all of them will make the cut.

“Others, we may find that in today’s technology it’s still a bit difficult to realize the task at the same cycle time as a human would do,” Robert says. “And that’s fine, right? You learn through that experience.”

The use cases that work at scale get deployed first. Then the cycle repeats with harder ones: “we train, we deploy, we scale.”

There’s plenty still to prove. A thousand robots over roughly six years across Schaeffler’s global network is not an instant flood flood. And Schaeffler is hedging its bets: CEO Klaus Rosenfeld told Reuters in May the company is working with around 45 humanoid robotics companies, including Agility Robotics , which it invested in back in 2024.

But Robert’s version of the factory of the future is refreshingly grounded. It’s not a dark factory with no lights and no people, but “one where autonomy and strong human-robot interactions coexist,” with robots on the repetitive work and the bottlenecks, and experienced workers freed up to fix processes and, eventually, ask a new question.

“What else can I do to maybe manage the fleet of robots instead of doing the task myself?”

Anyone who’s vibe-coding or using agents to get more done in the digital world knows that question very, very well. We’re just now on the cusp of being able to ask that question in the real world of physical AI.