Apptronik built hundreds of Apollo 2 humanoid robots and operated them as silent data collectors for more than a year before letting anyone outside the company see them. The Austin-based startup revealed that effort on Tuesday with the opening of Robot Park, a roughly 90,000-square-foot facility purpose-built to generate the training data that Google DeepMind's Gemini Robotics model needs to learn real-world logistics, manufacturing, and retail work.
That reframing matters more than the hardware. Apptronik CEO Jeff Cardenas, in comments carried by Reuters via The Star, framed the company as a two-output business: "We have a factory that produces robots, we also have a factory that produces data." The Robot Park unveiling is the first time investors and customers can compare those two production lines side by side.
The data side has been running longer than anyone outside the company knew. According to Humanoids Daily's account of the company's quiet website update, Apollo 2 has been in continuous internal use as a data-collection platform for more than a year. Apptronik has built "hundreds" of units, Cardenas told reporters, while declining to disclose how many are actually deployed at customer sites. That gap between units built and units working in the wild is the empirical question the company's narrative now has to answer.
The pipeline the facility anchors runs through a research partnership between Apptronik and Google DeepMind, whose destination model is Gemini Robotics, Google's robotics-specific AI. Every task the Apollo fleet performs inside Robot Park, from bin picking to pallet moving, is intended to produce labeled training data at a scale that simulation environments cannot match. Apptronik's own press materials and product page describe the new bipedal and wheeled Apollo 2 variants as both workers and sensors, designed to keep that data flowing as the fleet grows.
The commercial promise is straightforward: convert physical AI capital into a continuous stream of real-world robot training data, sell access to that pipeline through Gemini Robotics, and book the economics as a robotics company rather than a labeling service. The deployment timeline, with pilots running through 2026 and production versions targeted for 2027, sets the clock on that promise. The risk is equally straightforward: until customers, competitors, or independent benchmarks confirm the throughput, the data factory is a credible bet, not a measured one.
What to watch next is whether Apptronik discloses deployment numbers, signs a named commercial customer beyond pilots, or publishes any external evaluation of Gemini Robotics's improvement on the data the facility produces. Until one of those happens, Robot Park is a piece of AI plumbing with a hardware wrapper, and the interesting question is not how many robots Apptronik builds, but how many months of training data the facility can put on the model's plate.