Google DeepMindの新AIモデル「Gemini Robotics 2」がロボットの全身制御に対応Google DeepMind’s new AI model can control a robot’s entire body

Google DeepMind has released Gemini Robotics 2, an updated version of its robotics-focused AI model that the company says can "control entire humanoid robots." The upgrade matters because whole-body coordination has long been one of the hardest problems in humanoid robotics, and it signals that large AI developers are pushing their models beyond narrow tabletop tasks toward machines that can move and act more like people.

The headline change is the expansion of what the model can command. According to Google DeepMind, the previous version of Gemini Robotics focused primarily on controlling a humanoid robot's upper body, such as its arms and hands for manipulation tasks. Gemini Robotics 2 now supports what the company describes as "whole-body motions," spanning from a robot's feet to its fingertips. In practice, that appears to mean the model can coordinate locomotion, balance, and posture alongside the fine motor control needed to grasp and manipulate objects, rather than treating the upper and lower halves of a robot as separate problems.

That distinction is more significant than it may sound. Controlling a humanoid's legs and maintaining balance while the arms reach, lift, or carry involves managing shifting weight, contact forces, and stability in real time. Systems that handle only the upper body often assume the robot is stationary or externally supported. By extending control to the full body, Gemini Robotics 2 is positioned to tackle tasks that require a robot to walk, crouch, reach across a space, or stabilize itself while working, though the exact range of demonstrated behaviors and their reliability outside controlled settings remains to be seen.

Gemini Robotics belongs to a category increasingly known as vision-language-action, or VLA, models. These systems build on the multimodal foundations of large language models like Gemini, but instead of producing only text, they translate visual input and natural-language instructions into physical actions a robot can execute. The approach is meant to give robots more general-purpose flexibility, allowing them to interpret a spoken or written command and adapt to objects and environments they were not explicitly programmed for, rather than relying solely on hand-coded routines for each task.

The release fits into a broader and highly competitive push across the industry toward capable humanoid robots. Google DeepMind has previously worked with humanoid hardware makers, and companies such as Figure, Tesla with its Optimus program, Boston Dynamics, Apptronik, and others are all developing humanoid platforms or the software to run them. Chipmaker Nvidia has also invested heavily in robotics through simulation and foundation-model tools aimed at training robots more efficiently. The common thread is a bet that general-purpose AI models, trained on large and diverse datasets, can make robots easier to instruct and more adaptable than earlier generations built around task-specific programming.

For context, whole-body control has traditionally been the domain of specialized control engineering and reinforcement learning, often tuned for a single robot design. Folding that capability into a broader AI model that also handles perception and language is the more ambitious goal, because it could allow the same underlying system to be adapted across different robot bodies and tasks. Whether Gemini Robotics 2 generalizes cleanly across hardware, and how much task-specific tuning it still requires, are open questions that will likely become clearer as researchers and partners test it in real deployments.

As with earlier robotics announcements, it is worth treating demonstration results with some caution. Robotics systems that perform impressively in curated videos or lab conditions can behave differently in unstructured, real-world environments, where lighting, clutter, and unexpected obstacles complicate both perception and movement. Google DeepMind has not framed the model as a finished consumer product, and the practical timeline for such technology reaching factories, warehouses, or homes remains uncertain.

Even so, the move from upper-body manipulation to full-body control represents a meaningful step in how AI models are being applied to physical machines. It reflects a wider industry direction in which foundation models increasingly serve as the "brain" for robots, and it adds to the growing evidence that major AI labs see embodied systems as a key frontier alongside their work on text, image, and video generation.