On September 21, Boston Dynamics switched on a permanent training floor for Atlas inside Hyundai Motor Group Metaplant America, the Georgia plant outside Savannah. It is called the Robotics Metaplant Application Center, or RMAC, and the robots in it are working on one job: pulling car parts out of shipping packaging and arranging them in the order the line will install them.
RMAC is not a new name here. When we covered the production Atlas handstand on May 11, all 2026 Atlas output was already committed to RMAC and Google DeepMind, and the open question was whether RMAC would quietly slip into 2027. Later that month BD CEO Robert Playter described it as a data factory. It did not slip. Atlas now has a fixed address inside a car factory, and the details behind Playter’s label are far more specific than the label.
What the release commits to
The Boston Dynamics announcement lays out a three-step task ladder. Parts logistics and sequencing now. Component assembly by 2030. After that, repetitive and strenuous work such as lifting heavy loads.
The site is already outgrowing itself. Next year RMAC moves into a new building on the HMGMA campus at roughly ten times its current footprint. Also next year, Boston Dynamics plans to start exploring Atlas use cases outside automotive. It is in talks with existing Spot and Stretch customers in manufacturing, aerospace, semiconductors, logistics, food and beverage, and life sciences to scale data collection at their sites through 2027.
The release restates two numbers we covered in May: 25,000 Atlas units across Hyundai and Kia plants and a U.S. factory that can build 30,000 robots a year. Both are plans, not fleet counts. Hyundai’s July roadmap still puts actual deployment at the Georgia plant in 2028, starting with parts sequencing, with assembly work targeted for 2030.
The Robot Behavior Factory
The release is the scaffolding. The useful material came from Zack Jackowski, Boston Dynamics’ chief product and technology officer, in an interview with The Robot Report.
Earlier pilots, he said, were time-boxed, small, and run hands-on by engineers, meant to answer how a scaled deployment should be designed. RMAC is the answer: a permanent facility for data collection and application testing, run in coordination with the live auto plant. He named it a Robot Behavior Factory. Real parts, fixtures and workcells go in. So do Atlas robots and skilled pilots. What comes out is training data.
That data has three sources: teleoperation, reinforcement learning in simulation, and Universal Manipulation Interface handheld devices.
Jackowski also walked through why sequencing is harder than it sounds. Car parts must be handled directly, often by complex surfaces. Many have cosmetic or fragile areas that cannot be gripped. Large, awkward or fragile parts frequently need two hands. And each one has to be extracted from shipping packaging and inserted into high-density cubbies for the trip to the line.
Then the single most important sentence in the interview. The design assumptions behind production Atlas, he said, have been “very close to right on.” What the team is learning is the application: how to modify fixtures and packaging to make them easier for humanoids, how to collect data in production, how to build support processes around the robots.
Read that plainly. At this stage the thing being redesigned is the workstation, not the robot.
Against the cohort
On August 11 we lined up the three automotive humanoid programs and found the same first job at all of them: Figure at BMW, Apptronik at Mercedes, Atlas at Hyundai, every one assigned to kitting. MIT’s Ben Armstrong called it a million-dollar solution to a one-hundred-dollar problem.
RMAC doesn’t answer Armstrong. It changes the shape of the bet. BMW and Mercedes are running pilots to find out whether a humanoid can do the task. Hyundai has built a standing facility to manufacture the data that teaches it, and has already scheduled a building ten times larger.
Toyota is the other useful comparison. On September 18 it named 400,000 in-house robots across 60 plants, starting in 2028, framed as a way to capture skills held by 18,000 aging craftsmen. Same start year. Toyota’s pitch is preserving human skill. Boston Dynamics treats human demonstration as feedstock from day one.
Who feels it, and when
Line-side sequencers and kitters. This is the job RMAC is rehearsing, with real parts, in a working plant’s campus. In August we ranked it the most exposed role on the floor; RMAC gives that ranking a street address. Deployment is still dated 2028. The two-plus years between now and then are not a grace period. They are the data-collection runway.
Teleoperation pilots. Jackowski’s skilled pilots are a new job category, and Boston Dynamics has moved staff from Waltham, Massachusetts, to Savannah and is hiring across departments. The work product is demonstration data, and the behavior it trains is aimed at the kitter’s station. It belongs in the same ledger as the $50-an-hour contractors rating ChatGPT answers we covered on September 14: human judgment, captured piece by piece, converted into model capability. Here the judgment lives in the hands. How long the role lasts depends on how much demonstration the models still need, and Boston Dynamics’ own guidance runs data scale-up through 2027.
Fixture, packaging and process engineers. The least visible layer, and possibly the one that changes first. If fixtures and packaging are where the learning is, supplier packaging specs, line-side rack design and workcell layouts will start being written for humanoid hands. These roles gain work rather than lose it. The new line item is designing the workplace for the machine.
Material handling outside autos. The 2027 expansion takes data collection to aerospace, semiconductor, logistics, food and beverage and life-sciences customers. Kitting and material-handling roles in those sectors should expect the same capture process on their floors next year.
One more ledger entry. On August 25, Hyundai’s Korean union settled for a clause about sharing information after striking for consent over robot deployment, and the 2028 date did not move. That agreement was bargained in Korea. The robots are learning the job in Georgia.
External sources:
- Boston Dynamics, Boston Dynamics Opens Robotics Metaplant Application Center to Train Humanoid Robots for Manufacturing Tasks (September 21, 2026)
- The Robot Report, Boston Dynamics opens Metaplant Application Center to train Atlas humanoids (September 21, 2026)
- Hyundai Motor Group Newsroom, robotics strategy statement (July 2026)
- Humanoids Daily, Boston Dynamics Opens Atlas Training Center at Hyundai’s Georgia Factory (September 21, 2026)
