Google、「Gemini Robotics 2.0」を発表——器用さと安全性の向上を約束Google reveals Gemini Robotics 2.0, promising improved dexterity and safety

Google has introduced Gemini Robotics 2.0, the newest generation of its artificial intelligence system for controlling physical robots, with the company positioning improved dexterity and safety as the central advances. The release matters because it pushes Google's Gemini model family beyond software tasks such as text and image generation and into the physical world, an area where major AI developers are increasingly competing to build what are often called foundation models for robotics.

According to the announcement, Gemini Robotics 2 is made up of three separate models, but only one of them is currently available to the public. Google has not opened access to all three at launch, which suggests a staged rollout in which the remaining models may be reserved for select partners, additional testing, or a later general release. The company frames the suite as a step toward robots that can perform finer manipulation tasks and operate more safely around people and objects, though the practical extent of those improvements will likely become clearer only as developers put the released model to work.

The emphasis on dexterity is significant because fine motor control has long been one of the hardest problems in robotics. Tasks that people find trivial, such as folding fabric, threading a small part, or grasping an unfamiliar object without crushing it, require precise, adaptive responses to touch and shape. Framing dexterity as a headline feature indicates that Google is targeting these manipulation challenges rather than only navigation or simple pick-and-place operations. The parallel focus on safety appears aimed at the reality that robots increasingly share space with humans, where unpredictable movements or misjudged force can cause harm.

For context, Gemini is Google DeepMind's flagship multimodal model line, built to process text, images, audio, and other inputs together. Applying it to robotics extends that multimodal capability into an approach commonly described as vision-language-action, in which a model interprets what it sees, understands instructions phrased in natural language, and translates them into physical movement. Google's earlier robotics work established the groundwork for this direction, including systems designed for embodied reasoning, the ability to understand a scene and plan actions within it, as well as efforts to run models efficiently enough to operate on the robot itself rather than relying entirely on the cloud. Gemini Robotics 2.0 appears to continue that trajectory, although Google has not detailed how the three new models divide these roles.

The move also sits within a broader industry push toward general-purpose robotics AI. Companies such as Nvidia, which has promoted its GR00T platform for humanoid robots, along with startups like Figure and Physical Intelligence, and hardware efforts including Tesla's Optimus, are all pursuing the idea that a single, flexible AI model could power many different robot bodies and tasks. The central bet across these efforts is that the same scaling and generalization that improved language models can be applied to physical action, reducing the need to hand-program robots for each new task. Google's decision to iterate on a numbered version signals that it intends to treat robotics as an ongoing product line rather than a one-off research demonstration.

Several important details remain unclear from the initial announcement. It is not specified which of the three models is the publicly available one, what hardware it is intended to run on, or the terms under which developers can access it. The safety claims, in particular, are difficult to evaluate without independent testing, since robustness in controlled demonstrations does not always translate to reliable behavior in unstructured, real-world environments. Readers should therefore treat performance descriptions as the company's own characterizations until third-party evaluations emerge.

Even so, the release underscores how quickly large AI providers are extending their models from screens into machines. If Gemini Robotics 2.0 delivers meaningful gains in manipulation and safe operation, it could lower the barrier for building capable robots, but the staggered availability of its three models suggests Google is rolling out those capabilities cautiously rather than all at once.