Google DeepMind launches Gemini Robotics 2 for whole‑body robot control

Google DeepMind announced Gemini Robotics 2, a suite of three AI models – a vision‑language‑action (VLA) model, an embodied‑reasoning (ER 2) model, and an on‑device model – that together provide full‑body control for humanoid and other robots. The system can run locally on the robot, adapt to new hardware with under 200 training examples, and enable multiple robots to collaborate on tasks.

Demonstrations used the Apptronik Apollo 2 platform (with a 22‑degree‑of‑freedom SharpaWave hand), Franka‑type arms and other rigs. Reported success rates were about 68 % for tabletop picks, 76 % for shelf picks and 46 % for floor picks. Safety testing showed 97.9 % compliance with safety constraints and 93 % accuracy in detecting humans within one metre. Access is provided through the Gemini Live API, Google AI Studio and early‑access programs for trusted partners.

The release is positioned as a research preview with a waitlist; no commercial deployment is planned for the near term, but the technology is presented as a potential “intelligence layer” for future robot hardware.