Apptronik Builds a 90,000-Square-Foot Data Factory to Train Apollo 2 Humanoids

The hardest problem in humanoid robotics has never been building the robot. It has been teaching it to work. Apptronik just made its most concrete bet yet on how to solve that: a massive, purpose-built data factory in Austin, Texas, where fleets of its Apollo 2 humanoids clock in every day, not to ship products, but to generate the training data that will make future robots actually useful.

A Factory That Makes Data, Not Widgets

Apptronik announced the opening of the newly expanded Robot Park, its flagship data collection and training facility for humanoid robots in Austin, Texas, anchoring a growing global network of Robot Parks at customer and partner sites around the world, with plans to open new locations in more cities soon. Inside the nearly 90,000-square-foot facility, Apollo 2 robots in bipedal and wheeled configurations learn across an extensive array of customer use cases, performing tasks in logistics, manufacturing, and retail.

The facility operates seven days a week, with fleets of humanoid robots moving boxes, sorting items onto conveyor belts, and practicing logistics routines under the guidance of remote operators. Through a combination of teleoperation and autonomous execution, Apollo 2 robots continuously generate significant quantities of high-quality training data, which is used to train and refine the Gemini Robotics AI models that will prepare Apptronik's commercial fleet for real-world deployment.

The Google DeepMind Connection

As part of Apptronik's research partnership with Google DeepMind, the high-quality data collected by Apollo 2 helps to advance Gemini Robotics, Google DeepMind's foundational AI models for robotics. Gemini Robotics is a vision-language-action (VLA) model, meaning it takes in camera feeds and language instructions and outputs physical robot movements, rather than relying on hand-coded task scripts. Gemini Robotics can be specialized for complex embodiments such as the humanoid Apollo robot developed by Apptronik, with the goal of completing real-world tasks.

DeepMind has adapted Gemini Robotics On-Device, a lower-latency, locally-executed variant, to Apollo, enabling on-robot inference without cloud round-trips and requiring as few as 50 to 100 demonstrations to adapt to new tasks. That last number matters: the ability to teach a robot a new skill in under 100 demonstrations, rather than thousands, is what makes real-world deployment economically viable.