Google DeepMind Launches Gemini Robotics 2 for Physical AGI

Google DeepMind released the Gemini Robotics 2 model series on July 31, 2026, aiming to achieve physical AGI by providing intelligent whole-body control for humanoid robots. The suite consists of three components: Gemini Robotics ER 2 for high-level reasoning and multi-step planning, Gemini Robotics 2 as a vision-language-action model for physical control, and Gemini Robotics On-Device 2 for local execution and rapid adaptation.

The system allows robots to coordinate locomotion, arm motion, and object manipulation simultaneously. Using the ER 2 model, robots can track task progress and self-correct in real-time. Google demonstrated these capabilities through multi-robot collaboration featuring Apptronik's Apollo 2, Boston Dynamics' Spot, and a Franka F3 Duo arm. While the models can perform tasks like tidying shelves, sealing bags, and screwing in lightbulbs, some reports indicate robots remain slow and struggle with high-dexterity actions such as tying trash bags.

To manage the risks of frontier AI in physical spaces, Google introduced the ASIMOV-Agentic benchmark to test whether robots can detect human proximity, avoid collisions, and refuse unsafe commands. The models are currently available to developers via the Gemini API, Google AI Studio, and a private preview on the Gemini Enterprise Agent Platform.