Google DeepMind announced Gemini Robotics 2 on July 30, 2026. It is a set of AI models designed to be the brain inside the next generation of robots. The release includes three models that each handle a different part of what a robot needs to do: understanding instructions and moving, thinking through tasks, and running on the robot itself without an internet connection. DeepMind Blog

The first model, Gemini Robotics 2, takes what a robot sees and what a person tells it and turns that into physical movement. It can control a full humanoid robot — everything from its feet to its fingertips — as well as robots with two arms. In a demonstration, a humanoid robot using this system walked, crouched, reached, and picked up objects to clean up a messy room.

The second model, Gemini Robotics ER 2, gives robots the ability to understand the physical world around them, talk with people, and plan tasks that take several minutes to finish. ER 2 also introduces something new: multiple robots can now work together as a team on shared goals.

The third model, Gemini Robotics On-Device 2, is designed to run directly on a robot's own hardware. Its key feature is fast learning — it can figure out how to control a completely new robot body, with a different shape or different limbs, using only a few hours of data. Normally, teaching a robot to move well requires enormous amounts of practice data. This model cuts that down significantly.

This release builds on earlier work from DeepMind called Gemini Robotics, which first showed that the company's AI could turn its understanding of images and language into real physical action. That earlier version proved the concept. Gemini Robotics 2 expands it across reasoning, movement, and on-device execution.

The broader context here is how DeepMind has divided robot intelligence into separate pieces. Instead of one model trying to do everything, the company split the work: one model handles thinking and planning, another handles physical movement, and a third handles learning on the robot itself. In my view, this separation solves a real problem in robotics. When a single system tries to handle both high-level thinking and low-level movement at the same time, it tends to be slow and struggles when conditions change. By splitting these roles, each model can focus on what it does best.

The multi-robot teamwork in ER 2 deserves attention. Getting several robots to work together requires that each robot knows what the others are doing and that tasks get divided up sensibly. That is much harder than controlling a single robot. The design suggests that ER 2 handles the shared decision-making, while the movement model handles the physical work for each robot.

The on-device learning also solves a practical problem. New robot designs usually need huge amounts of data before they can be controlled reliably. Gemini Robotics On-Device 2 can adapt to a brand-new robot body with just a few hours of data, and it does so on the robot itself, with no need for a constant internet connection. That makes it easier for companies building robots with unusual designs.

Looking ahead, the combination of full-body control and on-device learning points toward a future where robots are more modular. Hardware makers can change the physical design of a robot without having to retrain the entire AI system from scratch. Having watched technology shifts push intelligence deeper into physical devices over the decades, this approach — an AI layer that works independently of the specific robot body — mirrors what happened in cloud computing and mobile phones, where separating software from specific hardware made everything easier to build and scale. A single AI controlling a humanoid from feet to fingertips, while a separate model handles long-term planning, sets up a division of labor that could speed up real-world robotics.