Google DeepMind Unifies Entire Humanoid Body under Gemini Robotics 2 and Takes Test to Apptronik, Franka, and Atlas
New model controls legs, torso, arms, and hands with 22 degrees of freedom under a single learned policy. Real deployment begins in the coming months with four partner manufacturers.
Google DeepMind announced on Thursday the Gemini Robotics 2, the first foundational model from the company capable of operating an entire humanoid under a single learned policy. The announcement, published on the official DeepMind blog and reported by Bloomberg on the same day, marks a shift from the previous generation, which needed to orchestrate control of arms, hands, and locomotion in separate models. Now, legs, torso, arms, and multi-fingered hands are coordinated by the same reasoning, with the model interpreting instructions in natural language and planning motor actions simultaneously.
The architecture was tested on a five-finger robotic hand called SharpaWave, mounted on Apptronik’s Apollo 2 platform, with 22 degrees of freedom. The team demonstrated task switching between the Apollo 2 and a bimanual platform F3 Duo from Franka, a movement that highlights DeepMind's interest in producing a portable foundational model that is not tied to a single manufacturer. Agile Robots, based in Munich with operations in China, and Boston Dynamics, whose agreement signed in May of this year provided DeepMind with a fleet of Atlas robots for integration with Gemini, are also partners in the program.
What Changes for the Enterprise Buyer
The technical gain is clear. A single model capable of reasoning about the entire body reduces the integration footprint for those looking to deploy a humanoid in a manufacturing line, a distribution center, or a hospital unit. Instead of hiring separate teams for navigation, manipulation, and vision, the client signs a contract with one of the OEMs and uses the DeepMind model as a unified cognitive layer. The learning curve for reprogramming the robot in light of a new SKU, a new workbench, or a new floor layout is expected to shrink, even though DeepMind acknowledges that the first field deployments, slated for the coming months, will serve to calibrate real limits outside the laboratory.
The geographical distribution of the partners tells a second story. Boston Dynamics, now under Hyundai, produces the new Atlas in Waltham, Massachusetts, and promised at CES to begin commercial shipments this year. Apptronik, based in Austin, has a pilot contract with Mercedes-Benz to operate the Apollo on a German assembly line. Franka Robotics has its manufacturing base in Bavaria, and Agile Robots operates between Germany and China. Readers who need to assess the geopolitical risk of an enterprise automation pipeline now have a concrete insight: cognitive software is from Mountain View, the steel chassis is divided among the United States, Germany, and China. Each of these jurisdictions brings its own export control regime, industrial data rules, and labor legislation.
Competition Heats Up
The launch puts DeepMind on a more direct path with Tesla's xAI Optimus, Figure AI, and Physical Intelligence, all of which have announced foundational humanoid models in the last twelve months. What sets Gemini Robotics 2 apart, according to Google’s own assessment, is its integration with the rest of the Gemini stack already used by CIOs for productivity and coding, which opens the possibility for the same executive agent to transition from text conversation to physical action via a robot with relative ease of handoff. It’s a promise, not yet a product, but an agent that communicates with the company’s CRM and then instructs the humanoid to fetch a part from inventory is no longer a theoretical scenario.
It’s important to consider the other side. Meta spent months trying to reposition its robotics strategy without announcing a reference product, and DeepMind itself has recently faced turbulence with the dissolution of the original AlphaFold team, a move the company attributed to a redirection of focus to Gemini. Analysts from Morgan Stanley and JPMorgan still project that the global humanoid market will not reach a significant revenue mass before 2029, a scenario that requires CIOs to exercise caution between enthusiasm and capital allocation.
What to Watch
The next test of the model will be a field deployment with Apptronik and Franka in the coming months. Two metrics warrant close monitoring: the reconfiguration rate for a new task within the same plant, currently the largest Achilles' heel of conventional collaborative robots, and the rate of human intervention per hour of operation, a clear indicator to separate demo from ready product. In parallel, DeepMind's decision to keep Gemini Robotics 2 outside the public API, currently restricted to program partners, suggests that Google intends to monetize through a slow OEM cycle rather than via an SDK open to startups. For enterprise buyers, this means that the entry point will continue to be the hardware supplier, not the cloud platform.