Gemini Robotics 2がロボットに全身インテリジェンスをもたらすGemini Robotics 2 brings whole body intelligence to robots

Google DeepMind has introduced Gemini Robotics 2, a new foundation model the company says gives robots "whole-body intelligence," allowing machines to coordinate movement across their entire physical form to complete complex, multi-step tasks. The announcement matters because most robotic manipulation systems to date have concentrated on arms and grippers, and extending learned intelligence to full-body coordination is widely viewed as a prerequisite for the general-purpose robots the industry has been chasing.

According to the DeepMind blog, Gemini Robotics 2 enables robots to plan and execute actions that require coordinating whole-body movements rather than isolated gestures. In practice, whole-body control means a robot must reason about balance, reach, base motion, and the interplay of multiple limbs at once — the kind of coordination people perform without thinking when they crouch to pick something up, brace against a surface, or shift their weight while carrying a load. Treating these as a single, unified control problem, rather than a sequence of separate manipulation steps, appears to be the central idea the release is built around.

The model is positioned as a successor to Google DeepMind's earlier Gemini Robotics work, which paired the Gemini family of multimodal models with robotic control. That lineage places Gemini Robotics 2 within the broader category of vision-language-action, or VLA, systems: models that take in camera images and natural-language instructions and output the motor commands a robot needs to act. This approach lets a robot connect the general world knowledge and reasoning of a large multimodal model to the specific physical demands of a task, and it is the same conceptual foundation underpinning much of the current wave of robotics research. A related strand of DeepMind's work has emphasized embodied reasoning, in which a model interprets a scene, breaks a goal into sub-steps, and adapts when conditions change.

Foundation models are attractive for robotics because they hold out the promise of generalization. Rather than programming a machine for one narrow task, researchers hope a single model trained on large and varied data can transfer skills to new objects, environments, and instructions with relatively little additional teaching. Whole-body control raises the difficulty of that goal considerably, since the model must produce coordinated, physically stable behavior in real time, where small errors can compound into a fall or a dropped object. The framing of Gemini Robotics 2 as "a significant step toward general-purpose robotic intelligence" reflects that ambition, though how well the capability holds up outside curated demonstrations is likely to be a key question as more details emerge.

The release lands amid intense competition in humanoid and general-purpose robotics. Nvidia has promoted its GR00T foundation-model effort and simulation tooling for training robots, while companies such as Figure, Tesla with its Optimus program, and Boston Dynamics with an electric Atlas have all pushed toward robots that can operate in human environments. Much of this activity converges on the same premise: that progress in large multimodal models can be transferred to physical machines, shortening the long-standing gap between what robots can understand and what they can reliably do. Whole-body coordination is a recurring theme across these programs, because tasks in warehouses, factories, and homes rarely stay within the fixed workspace of a stationary arm.

Several practical caveats remain. DeepMind's announcement centers on the model's capabilities rather than a specific consumer or commercial product, and the blog framing does not, on its own, establish which robot hardware platforms the system runs on, how broadly it will be made available, or the safety constraints that would govern real-world deployment. Robots that move their whole bodies near people introduce heightened safety and reliability requirements, and bridging the gap between laboratory results and dependable operation has historically been slow. Independent testing and third-party access will be important for judging how the model generalizes beyond the scenarios Google chooses to showcase.

For now, Gemini Robotics 2 is best understood as an incremental but meaningful research milestone in a fast-moving field, extending the vision-language-action paradigm from dexterous hands to coordinated, full-body behavior. Whether it accelerates the arrival of genuinely general-purpose robots will depend on details still to be disclosed, but it signals that Google DeepMind intends to remain a central player as foundation models increasingly move from screens into the physical world.