Atlas gets a factory curriculum at Hyundai’s Georgia plant

Boston Dynamics has opened the Robotics Metaplant Application Center (RMAC), a training facility for its electric Atlas humanoid inside Hyundai Motor Group Metaplant America near Savannah, Georgia. The first phase is operating now and will prepare robots for production work before a planned 2028 deployment at the plant.

RMAC is the first operating site tied to Hyundai and Boston Dynamics’ CES 2026 strategy. On-site training, a large internal order and direct access to Hyundai’s factory systems give Boston Dynamics a structured path from prototype testing to fleet operations.

Atlas gets a factory curriculum

Boston Dynamics CEO Robert Playter calls RMAC a “data factory” because each training run can improve both the robot and the company’s proprietary manufacturing dataset. Engineers can capture how Atlas lifts, turns, grasps, recovers from errors and works around people under factory conditions.

Atlas will begin with parts sequencing for the IONIQ 5 and IONIQ 9. The robot must retrieve components from bins and arrange them in the order required by the assembly line. The workflow has predefined parts, locations and handoffs, making it a practical starting point for measuring pick accuracy, cycle time and recovery behavior. It can also reduce repetitive lifting and awkward reaches for workers.

RMAC combines physical runs with simulation and operational data from Hyundai’s “Software-Defined Factory,” its term for connected production systems that collect data across the plant. Boston Dynamics says engineers can use those records to retrain robot control policies, test updates in simulation and return validated versions to the floor.

The rollout runs through 2030

Hyundai supplies the anchor demand

Hyundai has committed to deploying more than 25,000 Atlas robots across Hyundai and Kia manufacturing plants. The group is also building annual U.S. production capacity of 30,000 robots. The Atlas commitment equals roughly 83% of one year at that stated production rate, although the companies have yet to publish a detailed delivery schedule.

Hyundai Motor Group acquired a controlling stake in Boston Dynamics in 2021, making the planned fleet an internal anchor order. That demand can support manufacturing investment, supplier contracts and long-term software development before Boston Dynamics establishes broader commercial volume.

The group has divided the stack among three operations:

• Boston Dynamics develops the Atlas hardware platform, control software and training system.
• Hyundai Mobis manufactures the actuators that drive the robot’s joints.
• Hyundai Motor Group supplies the factories, production workflows and operational data.

Four-hour batteries, three-minute swaps

The Atlas at RMAC is the all-electric production model that replaced Boston Dynamics’ hydraulic research platform. The company has disclosed several specifications relevant to factory deployment:

• Power: Atlas has a four-hour battery life under typical use and can replace its own battery in less than three minutes. Autonomous swaps are designed to support continuous operation across shifts.
• Safety: An onboard system draws on autonomous-vehicle practices to detect people and vehicles. Atlas pauses when someone enters a configured safety radius, supporting layouts without fixed perimeter fencing. Each deployment still requires a site-specific risk assessment and safety controls.
• Availability: Boston Dynamics says all 2026 deployments are committed to RMAC and Google DeepMind, with additional customers planned for early 2027.

Every pick can become model data

Each physical run can generate synchronized records of sensor input, commanded motion, task outcomes and human intervention. A vision-language-action model can use such data to map camera input and task instructions to robot movements. Sequencing and assembly are especially useful because they require sustained planning, manipulation of multiple objects and repeated physical contact.

At Hyundai’s planned scale, Atlas could encounter variations in bins, parts, lighting, line speed and human behavior across multiple plants. Those variations can improve a model’s ability to handle unfamiliar conditions, provided Boston Dynamics captures failures and interventions as carefully as successful runs.

The resulting dataset will remain proprietary, and Boston Dynamics has yet to disclose its size, schema, update process or access rules. Google DeepMind’s 2026 allocation places production hardware with a major AI laboratory, but the announced terms leave model sharing and data ownership unspecified.

The customer list extends past Hyundai

Boston Dynamics is discussing Atlas deployments with existing Spot and Stretch customers in aerospace, semiconductors, logistics, food and beverage, and life sciences, according to launch reporting. RMAC’s larger planned building would give the company room to train and validate workflows outside automotive production.

Chief Product and Technology Officer Zack Jackowski also addressed SoftBank’s acquisition of the Robotics and AI Institute from Hyundai. The institute and Boston Dynamics have operated separately, he said, and the ownership change will not alter Atlas development.

“BD and RAI have been separate organizations with separate goals since inception, and the move has no impact on us,” Jackowski said.

Factory access shapes the humanoid race

Tesla has targeted its own manufacturing operations for Optimus, and Figure has deployed humanoids with BMW. Factories give these companies repeatable tasks, controlled environments and measurable outputs such as units per hour, error rates and intervention frequency.

Hyundai combines that factory access with ownership of the robot developer, actuator production and a planned fleet large enough to influence hardware design. The arrangement can shorten the cycle between observing a failure, updating the software and testing the change on the same workflow.

Scale still has unanswered numbers

The announcements provide a deployment roadmap but omit several figures needed to evaluate Atlas as a production system:

• Task performance: pick success, cycle time, recovery rate, uptime and human interventions per shift.
• Economics: unit price, service costs, maintenance intervals and expected payback period.
• Safety: applicable standards, validation methods and limits for operation around workers and vehicles.
• Learning system: the balance between simulation and physical training, onboard and cloud computation, and the process for deploying model updates.
• Fleet schedule: annual deliveries, allocation by plant and the number of robots assigned to each task.
• Workforce effects: changes to staffing, supervision, maintenance and intervention roles as deployments expand.