Google's new robot brain can watch, plan, and tag-team with other robots
Gemini Robotics ER 2 gives machines a smarter high-level 'thinker' that tracks progress on video and hands the actual moves off to a lower-level controller.
Key points
- Google DeepMind launched Gemini Robotics ER 2, a new 'embodied reasoning' model that acts as a high-level brain for robots.
- The model scores 57.4% on progress classification and 91.3% on moment-finding, spotting the exact video frame when a task step is done.
- It supports multi-robot collaboration, letting a wheeled robot and a humanoid share a task through common understanding.
- Developers can try it now through the Gemini API and Google AI Studio, with a private preview on the Gemini Enterprise Agent Platform.
- A demo pairs the model with Boston Dynamics' Spot robot to fetch objects on voice command.
Google DeepMind has released a new model that wants to be the thinking part of a robot's head. It is called Gemini Robotics ER 2, where ER stands for 'embodied reasoning', which is a fancy way of saying the model understands the physical world around it.
Here is the simple picture. The model watches through the robot's camera, listens to you, plans the steps, and then tells a separate lower-level controller how to actually move the arms and wheels. That lower-level controller is called a vision-language-action model, or VLA: software that turns a plan into specific motor movements.
Think of ER 2 as the manager and the VLA as the hands.
What can it actually do?
It can watch continuous video, track how a job is going, and switch to the next step at the right moment. That is the headline change from the previous version, ER 1.6.
Say a robot is pouring coffee. ER 2 is meant to see the exact frame where the cup is full and tell the arm to stop. Google's own testing puts it at 91.3% accuracy on this 'moment-finding' skill, with an average timing error under a second.
It also handles task-progress tracking, sorting each video frame into one of five buckets from 0% done to 100% done. That gives the robot a sense of 'am I nearly there, or did I mess up and need to try again?'
How is this different from other robot AI?
Speed and juggling. ER 2 plugs into the Gemini Live API, a two-way streaming connection built for low delay, so the robot does not freeze up every time it has to think.
It can also call outside tools mid-task, like Google Search, or a custom function a developer writes. Google showed a demo with Spot, the yellow four-legged robot from Boston Dynamics, fetching a bag of popcorn on a spoken request.
Another new trick: multiple robots working together. A wheeled robot and a humanoid can now share a semantic understanding of the job and hand tasks between each other. Google showed this with Apptronik's Apollo 2 humanoid and a Franka F3 Duo arm.
The numbers that matter
| Capability | ER 2 score | What it means |
|---|---|---|
| Progress classification | 57.4% | How well it tracks task progress on video |
| Moment-finding accuracy | 91.3% | Spotting the exact frame an event happens |
| Moment-finding timing | 0.96s average error | Sub-second reaction, fast enough for real robots |
Who can use it, and what does it cost?
Developers can access ER 2 today through the Gemini API and Google AI Studio, with a private preview on the Gemini Enterprise Agent Platform. Google has not published per-token robotics pricing, so hobbyists should expect standard Gemini API rates plus whatever hardware you point it at. Example code is on GitHub.
This is aimed at robotics developers and companies, not consumers. You cannot buy an ER 2 robot at Argos.
What about safety?
Google says ER 2 is its safest robotics model yet, with better scores on 'safety instruction following' and 'human proximity', meaning it is more careful about staying away from people and refusing unsafe commands. That is reassuring on paper. The real test will be homes, warehouses, and hospitals, where a spilled coffee is a nuisance but a misjudged arm swing is not.
One privacy note in plain words: any robot running this model is streaming video and audio to Google's servers to think. If these machines end up in your kitchen one day, that is worth remembering. For now, it is a developer tool, and the cameras are in labs.
