Google DeepMind unveiled Gemini Robotics 2 on July 30, a suite of three AI models that, for the first time, lets humanoid robots plan and execute movement across an entire body rather than just the arms and hands, the company said in its announcement. The system was demonstrated on Apptronik’s Apollo 2 humanoid, which DeepMind says can now walk, crouch, bend and manipulate objects while reasoning through multi-step tasks in real time.
Three models, one robot
The suite splits the work across three components. Gemini Robotics 2 itself is a vision-language-action model that turns what a robot sees and is told into full-body motor commands, covering everything “from feet to fingertips,” according to DeepMind. Gemini Robotics ER 2 is the higher-level planner: it breaks complex instructions into steps, communicates with people and, new in this release, coordinates several robots working on the same task. Gemini Robotics On-Device 2 runs locally on a robot’s own hardware and can adapt to a new robot body in a few hours, rather than the extensive retraining earlier versions needed.
DeepMind’s own benchmarks, published with the release, show wide variation by task type: whole-body manipulation scored roughly 46% to 76% success, gripper-based dexterity tasks scored about 74% to 90%, and fine multi-finger manipulation ranged from 32% to 92% — a spread the company says reflects how much harder some physical tasks remain for robots than others.
Betting on software, not hardware
The release continues Google’s strategy of supplying the “intelligence layer” for robots that other companies build, rather than building its own hardware — the same approach it took with the earlier Gemini Robotics ER 2 model. Beyond Apptronik, DeepMind said the new models have also been tested on hardware from Boston Dynamics, Franka and Agile Robots. Gemini Robotics ER 2 is now available through Google AI Studio and in private preview on the Gemini Enterprise Agent Platform, while the full VLA and on-device models remain limited to selected early-access partners.
The launch arrives as humanoid robots move from research demos toward real deployment, and as regulators pay closer attention to the hardware side of that shift — the FCC this year moved to bar new Chinese-made humanoid robots from US authorization over security concerns. For now, DeepMind’s push suggests the harder open problem in robotics may be teaching a machine to move at all, not just to talk.