Google bets Gemini becomes the default brains for humanoid robots

Google DeepMind is not building robots. It is betting it can supply the brains for everyone else's, at the moment Chinese manufacturers dominate humanoid shipments and hardware makers are racing to bolt their own AI onto their machines. Gemini Robotics 2, announced 30 July 2026, is DeepMind's clearest move yet to own that software layer before rivals lock in their own stacks.

The release is three models. Gemini Robotics 2 itself is the vision-language-action model, converting what a robot sees and is told into motor commands for full humanoid bodies and two-armed setups, dexterous hands included. Gemini Robotics ER 2 handles embodied reasoning: planning, following instructions, talking to people, and now coordinating several robots on the same task. Gemini Robotics On-Device 2 is the commercial hook — an efficient model that runs locally and adapts to a new robot body in hours using fewer than 200 examples, cutting the integration cost that has kept makers building in-house.

DeepMind says one checkpoint works across different robot bodies, including Apptronik's Apollo 2, Franka Duo, Dexmate, SO101 and Trossen platforms, and can plan over several minutes rather than single actions. Performance figures released alongside the launch show general whole-body manipulation succeeding 45.7-76.3% of the time, gripper tasks 74.2-89.6%, and multi-finger dexterity a wide 32-92%. DeepMind acknowledged the gap, saying "the multi-finger dexterous manipulation remains challenging."

Access is staged. Gemini Robotics ER 2 is available now in Google AI Studio and private preview; the VLA and on-device models are limited to early-access partners, meaning the commercial rollout is still gated rather than open.

The pitch to robot makers is clear: skip the years of building your own control stack and license Google's instead, with a model that already moves between five hardware platforms and adapts to a sixth in hours. That's a direct challenge to any manufacturer planning to differentiate on proprietary AI, and a gift to smaller makers who can't afford to build it themselves.

The risk is the dexterity numbers. A 32% floor on multi-finger tasks is not something customers deploying robots at scale can ignore, and DeepMind's own admission that this remains hard gives Chinese shipment leaders, who already have volume and cost advantages, room to close the software gap before Google's early-access partners turn into paying customers.