{"slug": "google-releases-gemini-robotics-2-model-that-gives-humanoids-full-body-control", "title": "Google Releases Gemini Robotics 2 Model That Gives Humanoids Full-Body Control And Lets Them Work In Teams", "summary": "Google DeepMind released Gemini Robotics 2, a family of three models that gives humanoid robots full-body control and enables teamwork, moving beyond tabletop demos to tasks like walking, crouching, and multi-step chores. The vision-language-action model drives full humanoids, the embodied reasoning model plans and tracks progress, and the on-device version runs without network connectivity. Success rates range from 30-45% for fine finger manipulation to 92% for unscrewing a lightbulb, while the embodied reasoning model scores 82.4% on success detection from video.", "body_md": "Google is lagging the frontier at general intelligence, but it’s producing some state-of-the-art AI models in specific verticals.\n\nRobotics is quickly turning into one of them. Google DeepMind has released [Gemini Robotics 2](https://officechai.com/ai/google-unveils-gemini-robotics-1-5-a-thinking-robot-brain-that-plans-before-it-acts/), the latest version of its family of robotics models, and this one moves past tabletop demos into something closer to what people actually picture when they hear the word humanoid. The model can now control a robot’s entire body, not just its arms, letting it walk across a room, crouch to pick something off the floor, reach up to a shelf, and carry out the kind of multi-step chores that used to need a person standing nearby with a joystick.\n\nThe update actually ships as three separate models rather than one. Gemini Robotics 2 is the vision-language-action model that converts what a robot sees and hears into actual motor commands, and it can drive full humanoids as well as simpler bi-arm setups. Gemini Robotics ER 2 is the embodied reasoning model, effectively the robot’s planning brain, responsible for breaking down instructions, tracking progress over a task that might run several minutes, and deciding when to call in another robot for help. Gemini Robotics On-Device 2 is the compact version built to run without a network connection, which matters a lot for warehouses and factory floors where connectivity can’t always be guaranteed.\n\n## Whole-body control finally arrives\n\nGoogle’s earlier robotics models were largely confined to the upper body, good enough for picking something up off a table but not much else. In the demo Google showed alongside this release, an Apptronik Apollo 2 robot is told to put a watering can into a bin on the bottom shelf. It walks to the table, picks up the can, crosses the room, and places it exactly where it was told, without anyone breaking the instruction down into smaller steps first.\n\nThe company is candid that speed remains a problem. The robots in Google’s own demos move at a pace that would frustrate anyone in a hurry, and the accuracy numbers back that up. On tasks like picking items off a table, floor, or shelf, Gemini Robotics 2 lands success rates ranging from the high-40s to the mid-70s in percentage terms, depending on the task, which tells you fairly plainly that this is progress, not a solved problem.\n\n## Dexterity is still the hard part\n\nWhere the numbers get genuinely humbling is finger-level manipulation. Tasks like tying a trash bag shut, zipping a ziplock bag, or using a dustpan sit in the 30 to 45 percent success range when Google’s model drives a five-fingered, 22-degree-of-freedom hand. Unscrewing a lightbulb, oddly, comes out far ahead of the pack at 92 percent, probably because it’s a repetitive rotational motion that’s easier to generalize than a task requiring fine sequential coordination. Simpler two-fingered gripper tasks on a Franka Duo arm fare much better, with precise insertion tasks hitting close to 90 percent.\n\nThe gap between gripper dexterity and multi-finger dexterity is really the story of where robotics stands right now. Grippers are a solved-enough problem to deploy commercially. Human-like hands, the kind needed for a robot to genuinely take over a chore in someone’s kitchen, are still very much a research project.\n\n## Reasoning and teamwork\n\nThe embodied reasoning side of the release leans on comparisons Google has published against rival models, including [Gemini Robotics ER 1.6](https://officechai.com/ai/google-unveils-gemini-robotics-1-5-a-thinking-robot-brain-that-plans-before-it-acts/), OpenAI’s GPT models, and Anthropic’s Opus. On success detection from video, a task that involves the robot judging for itself whether an action actually worked, ER 2 scores 82.4 percent, ahead of the field. On question answering benchmarks tied to embodied reasoning, it posts 78.5 percent against competitors clustered well below that.\n\nThe multi-robot collaboration piece is new territory for Google. ER 2 can now coordinate between different robots so that tasks too large or too varied for a single machine get split across a team, with the reasoning model acting as the coordinator relaying context between them. Google frames this as necessary for warehouse and logistics-style workflows where a single robot, however capable, will always be a bottleneck.\n\n## Running without the cloud\n\nGemini Robotics On-Device 2 gets less attention in Google’s announcement but might matter more for actual deployment. It’s built to run locally on hardware with no internet dependency, and Google says it can adapt to an entirely new robot body, different shape, different sensors, different degrees of freedom, in a matter of hours using fewer than 200 examples. That’s a meaningfully faster turnaround than retraining a model from scratch for every new piece of hardware, and it’s the kind of thing that could let smaller robotics startups building on [Apptronik’s Apollo](https://officechai.com/ai/top-humanoid-robot-companies-2026/) or similar platforms skip months of custom engineering.\n\n## Access\n\nGemini Robotics ER 2 is live now on Google AI Studio, with a private preview available on the Gemini Enterprise Agent Platform. The VLA and on-device models remain restricted to early-access partners for now, which tracks with how Google has rolled out every robotics model before this one. Full deployment to outside developers typically comes months after the initial announcement, once safety testing on physical hardware clears internal bars.\n\nGoogle is also introducing a new safety benchmark alongside the release, called ASIMOV-Agentic, aimed at testing whether a robot’s reasoning model will refuse an unsafe instruction from the action model, and whether it knows to pause and ask a human when it isn’t sure a task is even possible. Given that these models are increasingly being asked to control machines that can physically bump into people, that’s not a box Google can afford to leave unchecked.", "url": "https://wpnews.pro/news/google-releases-gemini-robotics-2-model-that-gives-humanoids-full-body-control", "canonical_source": "https://officechai.com/ai/google-releases-gemini-robotics-2-model-that-gives-humanoids-full-body-control-and-lets-them-work-in-teams/", "published_at": "2026-07-30 17:17:38+00:00", "updated_at": "2026-07-30 17:59:22.969602+00:00", "lang": "en", "topics": ["artificial-intelligence", "robotics", "ai-research", "computer-vision"], "entities": ["Google DeepMind", "Gemini Robotics 2", "Apptronik Apollo 2", "Gemini Robotics ER 2", "Gemini Robotics On-Device 2", "OpenAI", "Anthropic"], "alternates": {"html": "https://wpnews.pro/news/google-releases-gemini-robotics-2-model-that-gives-humanoids-full-body-control", "markdown": "https://wpnews.pro/news/google-releases-gemini-robotics-2-model-that-gives-humanoids-full-body-control.md", "text": "https://wpnews.pro/news/google-releases-gemini-robotics-2-model-that-gives-humanoids-full-body-control.txt", "jsonld": "https://wpnews.pro/news/google-releases-gemini-robotics-2-model-that-gives-humanoids-full-body-control.jsonld"}}