Google DeepMind unveils Gemini Robotics 2, giving robots 'whole body' intelligence

One of the toughest technical obstacles for humanoid robots has long been coordination: reaching for an object while maintaining balance and orienting the torso correctly requires managing dozens of joints simultaneously. Many robotic systems have historically handled this by breaking the task into pieces — stabilising first, then moving the arm — which tends to make movements slow and mechanical-looking.

Google DeepMind announced Gemini Robotics 2 on 30 July, a family of AI models aimed squarely at that problem. According to the company, the system gives robots the ability to plan and adjust their entire body, from torso to legs, as a single coordinated unit — letting a machine twist, lean and reach all at once.

In practice, that means a single integrated motion system that lets a humanoid robot walk, crouch, stretch and manipulate objects to clean up a cluttered room. Demonstration footage released by the company shows robots inserting a tape into a boombox, screwing in a lightbulb, and even tying up a garbage bag.

The system is built from three specialised models: a Vision-Language-Action model, an Embodied Reasoning model, and an on-device controller. That three-part structure separates high-level task planning — understanding what to do — from low-level motor control — figuring out how to physically do it — an architectural approach robotics researchers have long pursued.

Performance results vary considerably by task. The system achieved a 92% success rate on a task requiring precise whole-body control: unscrewing a lightbulb, demonstrating that fine motor skill and broader body coordination were being combined effectively.

But tests with a robot platform called Apollo 2 produced more mixed results: an 68.4% success rate picking objects up from a table, 45.7% from the floor, and 76.3% from a shelf.

The lower success rate for picking objects up off the floor illustrates a concrete challenge still facing the field: bending down requires shifting the body's centre of gravity while maintaining balance, perceiving the scene from a different visual angle, and applying precise grip in what is often a more constrained space for the hand to manoeuvre — all variables that must be managed simultaneously and in real time.

Still, researchers stress that this performance gap between tasks is expected, since each task demands a different combination of balance, perception and manipulation. The goal is not perfection at any single task but reasonably reliable general competence across a wide range of physical tasks.

Gemini Robotics 2 is part of Google's broader push into AI systems that operate in the physical world, a domain considered especially challenging because, unlike conversational AI models, it requires directly grappling with real-world uncertainty and physical constraints.

Experts say reliable, general-purpose humanoid robots for household or industrial settings are still likely years away from everyday use, but progress on this kind of whole-body coordination is seen as a sign that one of the field's most fundamental technical barriers — fragmented, sequential movement control — is beginning to be overcome.