"You should think of it as a 'Robot Behavior Factory,' which takes real parts, fixtures, workcells; adds Atlas robots and skilled pilots; and produces diverse, high-quality data for behavior training."
That is how Zack Jackowski, Boston Dynamics' chief product and technology officer, describes the Robotics Metaplant Application Center. It is a better description than "robot school," and a more honest one.
RMAC opened September 21 inside Hyundai Motor Group Metaplant America, the EV campus outside Savannah, Georgia. It began pilot operations in June and has now moved to full operation.
What Atlas is actually learning is narrow, and worth stating precisely: preparing logistics and sequencing automotive parts — arranging components in the exact order the line needs them.
What Atlas is actually learning is narrow, and worth stating precisely: preparing logistics and sequencing automotive parts — arranging components in the exact order the line needs them. Not welding. Not assembly. Not the tasks that dominate a body shop.
The schedule is similarly specific. Atlas is slated to take over parts sequencing at HMGMA from 2028. Component assembly follows by 2030. Only after that does the company talk about repetitive motions and heavy lifting.
So the headline number and the headline capability are on different clocks. Hyundai plans to deploy 25,000 Atlas units across Hyundai and Kia plants worldwide over the next few years.
It also intends to build a U.S. production facility capable of manufacturing 30,000 robots annually — a target The Robot Report places at 2028. No location has been disclosed.
A 25,000-unit fleet requires that factory to exist first. It does not yet.
Boston Dynamics is staffing up for it anyway. The company has moved some employees from its Waltham, Massachusetts headquarters to Savannah and is hiring across departments.
Why parts sequencing, of all things? Jackowski laid out the constraints, and they are the real content of this story. Car parts must be handled directly, which means interacting with complex surfaces. They carry cosmetic or fragile elements, so the robot has to know which areas are safe to grab. They are often large, awkwardly weighted, or fragile enough to require two hands. And they arrive in complex shipping packaging, then must be inserted into high-density cubbies for transport to the line.
That is a manipulation problem disguised as a logistics problem. It is also the part of factory work that is hardest to fixture away.
Which is why the most interesting quote in the evidence is not about the robot. It is about the building. "The biggest thing we've been finding is actually that the design assumptions we made for the production version of Atlas have been very close to right on," Jackowski said. What the team is learning instead is "how we might modify fixtures and packaging to make them easier for humanoids."
Read that twice. The adaptation is running in both directions. Hyundai is reshaping parts presentation around the robot, not just training the robot around the parts.
The training stack is a hybrid. Atlas learns from teleoperation, reinforcement learning in simulation, and Universal Manipulation Interface devices, according to Jackowski. The facility's job is to produce the real-world data those methods cannot generate on their own.
RMAC is explicitly the second act. "The previous trials were time-boxed deployments focused on lessons needed to design our scaled-up presence," Jackowski said, describing the earlier work as small-scale and very hands-on with the engineering team. "RMAC is the result of those lessons."
Next year, operations move into a separate building on the HMGMA site, expanding the center to roughly ten times its current size. That is the milestone to watch. A tenfold expansion is a claim about data volume, and data volume is the actual bottleneck.
The expansion also sets up the pivot. Boston Dynamics says it will begin exploring Atlas use cases in other industries next year, and it is already in talks with existing Spot and Stretch customers in manufacturing, aerospace, semiconductors, logistics, food and beverage, and life sciences to scale data collection in 2027.
That is where this stops being a captive automation project and starts being a product. Jackowski is candid about why the Spot base matters: Spot data builds understanding of industrial facilities, which he calls a very rare type of data. But the bigger value, he says, is the institutional learning customers accumulate on how to deploy, integrate, and manage robot fleets. "Our customers that started with Spot are in a much stronger position."
Fleet management is the unglamorous capability that decides whether 25,000 units is a deployment or a warehouse full of idle hardware. Nothing in the evidence suggests that problem is solved. It suggests Hyundai is buying time on it.
One corporate note worth clearing: SoftBank's acquisition of the Robotics and AI Institute from Hyundai will have no effect on Atlas development, Jackowski said, calling the two organizations separate since inception.
The competitive frame is set by Toyota. The Toyota Group aims to deploy close to 400,000 robots, with 150,000 units across Toyota Motor Corp. and the Toyota Group starting in 2028 and 250,000 more at suppliers — a mix of replacements and new units including humanoids. Reuters reported the automation effort could cost up to $6.4 billion per year.
Toyota's number is larger because it counts everything. Hyundai's 25,000 counts only Atlas. Those are not comparable figures, and treating them as such is the most common error available here.
There is a second thing to keep straight. What Boston Dynamics demonstrated at CES 2026 was a production-focused Atlas designed for industrial tasks, not a stunt reel. What remains unclear is whether Atlas's full 2026 capacity is committed to Hyundai factories, how far deployments extend beyond tightly controlled settings, and what the service obligations look like once these machines are running real shifts.
The reason that matters is the control layer. Once a robot handles parts and triggers other equipment, the vendor that owns the update pipeline owns safety patches, feature timing, and which third parties can touch the environment. Inside Hyundai's own plants, that is vertical integration and largely invisible. Sold into aerospace or semiconductors, it becomes the negotiating position.
For now, the concrete scoreboard is small and checkable. Does parts sequencing actually start at HMGMA in 2028? Does the 30,000-a-year factory get a location and a groundbreaking? Does the tenfold expansion happen on schedule next year?
Jackowski's own framing sets the bar. If Atlas's design assumptions were "very close to right on," the next two years will show it in cycle times and uptime, not in press releases. A Robot Behavior Factory is a reasonable name for the facility. It is also an admission that behavior, not hardware, is the thing being manufactured.