On July 30, 2026, Google DeepMind announced Gemini Robotics 2, a family of three models aimed at moving robot foundation models from arm-and-gripper manipulation to control of an entire body. Gemini Robotics 2 is the vision-language-action model that converts vision and language input into motor control. Gemini Robotics ER 2 is the embodied reasoning model that plans multi-step tasks, communicates with people, and coordinates multiple robots working in a shared space. Gemini Robotics On-Device 2 runs locally on the robot and adapts to entirely new robot bodies from a few hours of data and, DeepMind says, fewer than 200 examples.
The headline capability is whole-body control of humanoids: walking, crouching, bending, and manipulating objects treated as one control problem rather than a locomotion stack with a manipulation stack bolted on top. DeepMind reports success rates of 45.7 percent to 76.3 percent across whole-body manipulation picking tasks, 32 percent to 92 percent on multi-finger dexterity tasks, and 74.2 percent to 89.6 percent on gripper dexterity tasks. The spread in those numbers is the honest part of the announcement: some tasks are close to reliable and some are not.
The demonstrations run on a range of hardware, including Apptronik’s Apollo 2 humanoid fitted with SharpaWave and Inspire hands, a Franka Duo with a Robotiq gripper, and Dexmate, SO101, and Trossen platforms. That breadth matters more than any single number, because the persistent problem in robot learning has been that policies do not transfer across bodies. Availability is staged: Gemini Robotics ER 2 is offered through Google AI Studio and a private preview, while the VLA and on-device models go to early-access partners only.
For anyone evaluating humanoid robots as a real deployment option rather than a demo reel, this is the layer that decides the timeline. Hardware has not been the binding constraint for a while; general-purpose control that survives contact with an unstructured environment has been. Success rates in the 45 to 76 percent range on whole-body picking are not production numbers, but they are being reported openly on named hardware, which is a considerable change from the video-only claims that dominated humanoid robotics a year earlier.