Gemini Robotics 2: Google DeepMind Teaches AI to Control Humanoid Robots

In July 2026, Google DeepMind unveiled Gemini Robotics 2, an updated version of its AI model for controlling humanoid robots in the physical world. Wired calls it "a significant leap toward physical AGI": artificial intelligence that acts not only in the digital realm but also physically interacts with objects and people.

What can Gemini Robotics 2 do?

The model's core idea is a single architecture for three tasks: perception (computer vision and spatial understanding), interpretation (processing natural-language instructions) and action (generating commands for motor mechanisms). Traditional robotic systems handle these three tasks with separate modules; Gemini Robotics 2 combines them in one multimodal model.

This means an operator can give a robot a voice command, such as "pick up the red object on the table," and the model builds the whole chain on its own, from visually recognizing the object to executing a precise grasping motion.

• Developer: Google DeepMind
• Platform: humanoid robots
• Architecture: a single multimodal model (vision + language + motor control)
• Positioning: the next step toward "physical AGI"

Its predecessor, the first-generation Gemini Robotics, laid the groundwork for the concept of "embodied AI": systems that can act in a three-dimensional physical environment rather than only process text. Gemini Robotics 2 builds on that concept, widening the range of tasks it can handle and improving control precision on complex humanoid platforms.

Why is physical AI riskier than digital AI?

Moving AI from the digital world into the physical one creates a fundamentally different risk profile. A language model can instantly correct a bad answer; a robot cannot. A wrong movement of a manipulator means a broken object, a spilled substance or real danger to people nearby.

"Bringing AI into the real world carries risks,"

Wired states, listing the technical and ethical challenges that come with Gemini Robotics 2.

Among the key problems are how the model behaves in situations that never appeared in its training data, who is legally liable when it makes a mistake, and how to keep physical interaction with people safe. These questions make physical AGI a regulatory challenge as well as an engineering one.

Google DeepMind is not the only company in this race. Tesla is developing its Optimus humanoid robot, Figure AI has raised more than $675 million in funding, and Boston Dynamics is pushing forward with Atlas. But Gemini Robotics 2 arrives with a fundamental difference: it is built on a multimodal Gemini model rather than on a loose set of control modules with weak links between them.

What is happening in the robotics market

The humanoid robot market is on the verge of large-scale growth. Goldman Sachs forecasts that it could exceed $150 billion by 2035, provided the industry solves key problems of reliability and safety. Competition is shifting from mechanical precision to intelligence: the ability to understand unstructured tasks, adapt to changing conditions and work alongside people without physical barriers.

Gemini Robotics 2 targets exactly this segment. It is a system with "general" intelligence in a physical body, not a narrowly programmed industrial manipulator that performs one movement perfectly but is helpless in a non-standard situation.

What this means

The launch of Gemini Robotics 2 confirms that leading AI labs see physical robotics as the next technological frontier. For business, it opens up new applications in logistics, manufacturing and service industries. For regulators, it poses an urgent task: developing safety standards for AI systems whose actions have irreversible physical consequences.