Gemini Robotics ER 2 発表：ロボット向け映像理解と協調制御が進化Introducing Gemini Robotics ER 2

Google has announced Gemini Robotics ER 2, an updated version of its embodied reasoning model built to give robots stronger perception, planning, and coordination abilities. The company describes the release as a step change in three areas that matter for real-world deployment: video understanding, tool orchestration, and collaboration among multiple robots. Because much of robotics still struggles to bridge the gap between polished laboratory demonstrations and reliable operation in messy environments, advances in how a model interprets its surroundings and sequences actions are meaningful for the field.

The "ER" in the name refers to embodied reasoning, a strand of the Gemini Robotics program developed with Google DeepMind that focuses on understanding physical space and planning what to do, rather than directly generating low-level motor commands. In broad terms, the Gemini Robotics family has been presented in two complementary parts. One is a vision-language-action approach that turns instructions and camera input into robot movements. The other, the ER line, concentrates on higher-level tasks such as spatial reasoning, scene interpretation, and deciding how a job should be broken down and executed. ER 2 appears to build on the first-generation ER model along these same lines, extending its reasoning rather than replacing the underlying design.

The emphasis on video understanding suggests improvements in reasoning across time rather than from single still images. Robotic tasks often depend on tracking how objects and people move, noticing when a step has completed, and reacting to changes as they unfold. Stronger temporal understanding is likely to help a robot follow multi-step procedures, recover from errors, and interpret dynamic scenes more reliably, though the exact scope of the gains will become clearer as developers test the model.

Tool orchestration reflects a trend already visible in large language models, where systems increasingly call external tools and services to accomplish goals rather than relying only on what they can do internally. Applied to robotics, this points to a model that can coordinate specialized functions, such as calling a perception module, querying a database, or invoking a motion-planning component, and stitch them together to complete a task. This kind of agentic behavior is central to moving from narrow, hard-coded routines toward more flexible systems that can adapt their approach depending on the situation.

Multi-robot collaboration is arguably the most notable of the three headline capabilities. Coordinating several machines to work on a shared task introduces challenges around communication, task allocation, and avoiding conflicts that a single robot does not face. Improvements here are relevant to settings such as warehouses, manufacturing lines, and logistics operations, where fleets of robots increasingly operate in the same space. Google positions these advances as steps toward more practical and capable robotic applications, though as with any vendor announcement, independent evaluation will be needed to judge how the model performs outside controlled conditions.

The release lands amid intense industry interest in what is often called physical AI or embodied AI. NVIDIA has promoted its foundation-model efforts for humanoid robots, and companies including Figure, Physical Intelligence, and Tesla have pursued their own approaches to general-purpose robot control. Against this backdrop, Google's strategy of pairing an action-oriented model with a reasoning-focused counterpart reflects a wider bet that robotic intelligence will come from combining perception, language understanding, and planning rather than from a single monolithic controller.

For readers less familiar with the underlying concepts, it helps to note that vision-language models form the foundation of much of this work. These systems are trained to associate images and text, which lets them describe scenes, answer questions about what they see, and follow written instructions. Extending that capability into the physical world requires grounding the model's outputs in real actions and constraints, which is where embodied reasoning becomes important.

At the time of the announcement, the most consequential details for practitioners will be how the model is made available, which robots and platforms it supports, and how it measures against prior versions and competing systems. Google frames ER 2 as a meaningful advance for real-world robotics, and while the described capabilities are substantial on paper, their practical impact will depend on testing across the varied and unpredictable conditions that define working environments.