Google DeepMind just made a big move in robotics.
On Tuesday, the company launched a new version of its AI model called Gemini Robotics On-Device. What’s special? It can run directly on robots — no internet needed.
This means robots can now understand and perform tasks all on their own, without relying on cloud servers.
Smarter Robots, No Wi-Fi Required
This new model builds on the earlier version of Gemini Robotics that Google released back in March. But instead of running in the cloud, Gemini Robotics On-Device works right on the robot’s own hardware.
That’s a huge shift. It means robots can work in places without internet access — like warehouses, homes, or even disaster zones — and still understand what they need to do.
What Can It Do?
Google showed off some real-world demos to highlight the model’s power.
In these demos, the robots:
- Unzipped bags
- Folded clothes
- Handled objects they’d never seen before
- Worked on an industrial belt to assemble parts
What’s cool is that the model started off trained on a specific robot type (called ALOHA). But it was later adapted to work on completely different machines — including:
- The Franka FR3, a robot with two arms
- The Apollo humanoid robot made by Apptronik
And it worked just as well — even in brand-new situations.
Nearly Cloud-Level Performance, But Local
Google says this new on-device model performs almost as well as its cloud-based version.
They didn’t name names, but according to Google, Gemini outperforms other local AI models in benchmark tests.
That’s impressive, considering it runs entirely offline.
Easy for Developers, Too
Google isn’t just releasing the AI model. It’s also giving developers the tools to make the most of it.
They’re rolling out the Gemini Robotics SDK, which helps developers train robots faster and easier. Using this toolkit, developers can:
- Show the robot just 50 to 100 task examples
- Use a virtual environment called MuJoCo to test everything before going real-world
In short: less code, more doing.
Everyone’s Racing Into Robotics
Google isn’t the only company working on smarter robots. Others are jumping in too:
- Nvidia is building a platform to train general-purpose robot models
- Hugging Face is working on open-source AI for robots, plus building robots themselves
- RLWRLD, a startup from South Korea backed by Mirae Asset, is also creating foundation models for robotic systems
It’s clear: the future of AI isn’t just on screens — it’s moving, grabbing, and building right in front of us.
Why This Changes Everything
Robots used to need constant internet access to “think.” But that’s not always practical. What if you lose Wi-Fi? What if you’re in a remote area?
With Gemini Robotics On-Device, robots don’t need to phone home anymore. They can act instantly, securely, and without delay.
That opens the door for:
- Faster responses
- Better privacy
- Lower operating costs
- More reliable performance
Whether it’s helping in factories, homes, or hospitals, this kind of local AI makes robots more useful and more independent.