Apollo 2 pairs a factory ready humanoid with a 90,000 square foot Austin training center and Google DeepMind's Gemini Robotics model. The thesis Apptronik is now selling: real customer work, not demos, is what trains the next generation of humanoid robots.
Apptronik built its newest humanoid, Apollo 2, to do two things at once. It is designed to take on real work in warehouses, factories, and retail backrooms. It is also designed to be a data collection machine, a sensor-laden workhorse that streams its daily tasks into Google DeepMind's Gemini Robotics foundation model, so that each unit on a customer floor makes every other unit a little smarter.
That coupling, a humanoid on the job, a dedicated training facility, and a foundation-model partner consuming the output, is what Apptronik is actually selling. CEO Jeff Cardenas calls it a "continuous learning loop." Whether that loop holds in real customer environments is the open question the company now needs to answer.
Apollo 2 comes in two configurations: bipedal, for tasks that demand a human-like form factor, and a wheeled-base variant for industrial settings where stability and payload matter more than legs. Both share the same sensor stack and onboard compute aimed at teleoperation, in which a human pilot remotely guides the robot through tasks the model has not yet mastered, and at autonomous execution, where the model takes over. Every teleoperated and autonomous session becomes training data.
The training infrastructure sits in Austin, where Apptronik has expanded Robot Park, its flagship humanoid data collection and training facility. Forbes pegs the site at roughly 90,000 square feet, framing it as a "humanoid data factory." It joins a network of smaller Robot Parks already running at customer and partner sites worldwide, where Apptronik staff and partner teleoperators drive robots through the messy, repetitive tasks real workplaces demand.
The model that consumes all of this data is Gemini Robotics, Google DeepMind's foundation model built specifically for physical robotics. Apptronik's research partnership with DeepMind is not new. Forbes coverage from late 2025 already showed Apollo executing home and warehouse tasks under the partnership. Apollo 2 is the platformized next step: a hardware line, a controlled training facility, and a capital plan built around keeping the loop running.
The capital is real. Apptronik closed an oversubscribed $935 million Series A earlier this year, with Google participating. That sum is what makes the multi-site data factory and the hardware refresh possible. It is also what Apptronik will be judged against: not the demo reel, but whether the flywheel produces measurably better robots quarter over quarter.
The honest gap is worth naming. Every data collection site referenced in Apptronik's launch is one the company itself controls: its own Robot Parks, plus customer and partner sites where Apptronik staff and teleoperators run the sessions. No independent benchmarks, no third-party performance figures, and no paid-customer shipment counts appear in the public material. Cardenas's framing of "real environments, real tasks" is a company-stated thesis, not an externally validated one. Until an outside operator reports measurable throughput gains from an Apollo 2 fleet on its own floor, the flywheel remains aspirational.
There is also the matter of who actually owns the model side. Google DeepMind is not betting on Apptronik alone. Per CNBC reporting from March 2026, Google has also struck a parallel Gemini Robotics partnership with Agile Robots, the Munich-based industrial automation firm. Apptronik is one humanoid robot builder inside a multi-vendor Gemini Robotics strategy, not the sole beneficiary of Google's robotics AI push. Cardenas's competitive position depends on Apptronik's hardware, data throughput, and customer footprint outpacing other Gemini Robotics partners working on the same model.
What to watch next: a first named customer willing to publish fleet-level productivity data from an Apollo 2 deployment, and any DeepMind milestone that ties Gemini Robotics model improvements to a specific humanoid data partner rather than to the model as a whole. Until those arrive, Apptronik's pitch is the cleanest articulation of the data-flywheel thesis in commercial humanoids, and the largest unverified one.