Google's Gemini Robotics 2 wants to build a robot that can do anything you ask
A new family of AI models from Google DeepMind pushes physical robots closer to general-purpose helpers, with better dexterity, multi-robot teamwork, and one model developers can try right now.
Key points
- Google DeepMind launched Gemini Robotics 2, a new family of AI models for physical robots, in 2025.
- The release includes three sub-models, one of which is publicly available to developers today via the Gemini Live API.
- Gemini Robotics ER 2 is an upgraded "embodied reasoning" model that DeepMind says is a significant step beyond its previous 1.6 version.
- The goal is a "generalist robot" that can respond to any spoken or typed instruction, a concept Google's scientists call "physical AGI" (artificial general intelligence).
- The new system can control an entire humanoid robot body, including complex multi-finger hands.
For years, the viral videos have followed the same script: a sleek robot does a backflip, holds a pose, then stops. Impressive. Also pointless. Those machines were built to do exactly one thing, and nothing else.
Google wants to change that completely.
What did Google actually announce?
Gemini Robotics 2 is a new set of AI models, meaning software brains trained on vast amounts of data, designed to run inside physical robots. Where earlier systems could handle narrow, pre-programmed moves, this release aims for something broader.
The headline goal is what DeepMind researchers call "physical AGI", shorthand for artificial general intelligence applied to the real world. The idea is simple to state and hard to build: you tell a robot what to do, and it figures out how. No special programming for each new task. No rigid script.
Three sub-models sit inside the Gemini Robotics 2 family. The most accessible right now is Gemini Robotics ER 2, the "ER" standing for embodied reasoning, which means the ability to understand a physical environment and act within it. Developers can connect to it today through the Gemini Live API, a programming interface that lets software teams build and test robot behaviours.
As first reported by Ars Technica AI, DeepMind describes ER 2 as a significant leap over the previous Gemini Robotics 1.6 model.
What can these robots actually do now?
Three improvements stand out from Google's announcement.
First, the system can handle whole-body control, meaning it coordinates a full humanoid robot, head to toe, including hands with multiple moving fingers. That kind of dexterity has been a stubborn bottleneck.
Second, the robots can continuously read a changing environment rather than relying on a fixed snapshot. A box moves; the robot adjusts.
Third, multiple robots can now collaborate on a single task, passing objects between them or dividing up steps of a job.
| Capability | Previous version | Gemini Robotics 2 |
|---|---|---|
| Whole-body humanoid control | Limited | Yes |
| Multi-robot collaboration | No | Yes |
| Continuous environment analysis | Partial | Improved |
| Developer API access | No | Yes (ER 2 today) |
Should ordinary people care yet?
Not quite yet, but the direction matters. Right now, Gemini Robotics 2 is a developer tool, something for labs and companies building robots, not a product you can buy.
What Google is really doing is closing the gap between a robot that performs a rehearsed trick and one that could, one day, help in a warehouse, a kitchen, or a care home. Each of these releases narrows that gap a little further.
The moment robots can genuinely follow open-ended instructions reliably, the economic pressure to deploy them will be significant. That is worth watching, even if your household robot is still years away.
Common questions
What is "physical AGI" and why does Google use that phrase?
AGI, artificial general intelligence, refers to an AI system capable of handling any task a human can, rather than just one specific job. Google's scientists use "physical AGI" to describe that same flexibility applied to robots operating in the real world.
Can developers use Gemini Robotics 2 today?
One of the three sub-models, Gemini Robotics ER 2, is available now through the Gemini Live API. The other models in the family are not yet publicly accessible.
