Google DeepMind launches Gemini Robotics 2 with full-body robot control

Google DeepMind has announced Gemini Robotics 2, an updated model for controlling robots. The company says it is the first to provide full control over a machine's entire body and let it reason about every movement in real time.

What changed in Gemini Robotics 2

The key new feature of Gemini Robotics 2 is full-body control. Traditional robot control systems program individual movements in advance: the arm lifts according to one algorithm, the leg takes a step according to another. This approach works in predictable conditions but fails at the slightest change in the environment.

According to Google DeepMind, Gemini Robotics 2 takes a different approach. The model reasons about each movement as it happens, much as a person does when walking on uneven ground or grasping an object of unfamiliar shape. The system coordinates all parts of the body at the same time rather than controlling them one after another.

• Developer: Google DeepMind
• Model: Gemini Robotics 2
• Key feature: full-body control
• Principle: reasoning about every movement in real time
• Result: a wide range of physical tasks previously out of reach for robots

Why it matters for robotics

Full-body control is one of the key unsolved problems in the field. Most industrial robots handle repetitive movements on production lines well, but they get lost in nonstandard situations: a dropped object, an unfamiliar container, a changed route.

Humanoid robots, which Tesla, Figure AI, Boston Dynamics and other companies are actively developing, require exactly this capability: dynamically coordinating the whole body in real time. If Gemini Robotics 2 solves this at the foundation-model level, it could speed up teaching robots new tasks without separately programming every movement.

"Gemini

Robotics 2 allows robots to reason about every movement, opening up a wide range of tasks," says Google DeepMind's official statement, as quoted by The Robot Report.

Context: Gemini in physical AI

Google DeepMind has steadily expanded the use of its multimodal models in robotics. The first version of Gemini Robotics focused on perception and basic manipulation. Judging by the description, Gemini Robotics 2 takes the next step, moving from recognition to active control of the body.

Physical AI became one of the most competitive areas in the industry in 2025–2026. OpenAI, Google, Meta*, and dozens of startups are increasing their investment in robots that can operate in unstructured environments: warehouses, hospitals, and homes.

What this means

If Gemini Robotics 2 really does provide general reasoning about body movements, it is a serious bid by Google DeepMind for leadership in Physical AI. For the industry, it means that foundation models could take over tasks that previously required specialized programming for each type of robot and each specific task.

*Meta has been designated an extremist organization and banned in Russia.