
Google DeepMind just dropped something kinda major: a new AI model called Gemini Robotics On-Device.
And yep, the name says it all—it runs directly on robots, allowing them perform tasks locally without needing an internet connection.
Here’s what actually makes it cool:
It’s smarter than your average offline bot: This model is basically a local version of DeepMind’s earlier Gemini Robotics release from March—and according to Google, it performs almost as well as the cloud version.
It runs fully offline-- so the robots can fold clothes or unzip bags without calling back to a data center.
You can control it using plain ol’ natural language: Just tell the robot what to do—in English (or any supported language)—and it listens.
But the best part? It actually does useful stuff.
In demos, Google showed robots doing tasks like unzipping bags and folding clothes. But it goes deeper—this model also learned to adapt to different robots and new, never-seen-before scenarios, like:
The bi-arm Franka FR3, which handled unexpected industrial assembly tasks.
Apptronik’s Apollo humanoid robot, which also got in on the action.
Oh—and there's a developer-friendly SDK, too.
Google released a brand-new Gemini Robotics SDK, letting devs train robots by showing them just 50 to 100 demonstrations using the MuJoCo physics simulator.
And in case you’re wondering—Google’s not alone in this robot race:
Nvidia is building out AI foundations for humanoid robots.
Hugging Face is diving into robotics datasets and even building their own bots.
RLWRLD (a Korean startup backed by Mirae Asset) is also deep into creating foundational models for robots.
So yeah... the race is officially on—and now the bots don’t even need Wi-Fi to keep up.