Google DeepMind announced Gemini Robotics 2 on 30 July 2026, an upgrade to its embodied reasoning model family, made available to developers through the Gemini API and Google AI Studio.
The central change is scope of control. Earlier models in this line largely governed a robot's hands and arms. Gemini Robotics 2 controls the whole body, which means a humanoid can walk to a location, crouch, reach, stretch and manipulate an object as one continuous action rather than as separate systems handing off to each other.
The model also handles multi-robot coordination, allowing several machines to work on the same task together, and adds real-time video understanding and task progress tracking, so a robot can assess how far through a job it is rather than simply executing a fixed sequence.
Two practical details stand out. The model can run locally on the robot rather than requiring a connection to remote infrastructure, which matters in environments with unreliable networks and for tasks where latency is a safety question. And DeepMind reports it can be adapted to an entirely new robot body in a matter of hours, rather than the retraining effort such a transfer has historically required.
It is built on Gemini 3.5 Flash, with improvements in spatial reasoning and video understanding carried over from the language model line.
The significance is less any single capability than who is now building in this layer. General-purpose embodied models are being released by large AI labs, by robot manufacturers, and by open-source projects at the same time, on overlapping timelines.
Share:
Humanoid Robots Set Records at the 2026 World Humanoid Robot Games