Google DeepMind wants teaching a robot to feel like texting a friend. Its new Gemini 4 Hydrogen model now plugs into robot arms through Google's Antigravity coding app, so users can train an arm in a virtual practice world and then send it to do real jobs, all by chatting in plain English.
Gemini 4 Hydrogen: Key Takeaways
- A new robotics plugin for Google Antigravity lets Gemini 4 Hydrogen control supported robot arms
- Users practice tasks in MuJoCo, a physics simulator, before running them in the real world
- Google DeepMind says Hydrogen finishes tasks 2.5 times faster than its older 3.8 Flash model
- The supported arms and a wider rollout date have not been named
Google DeepMind shared the update on Tuesday. The flow has three steps: connect a supported arm, rehearse the task in MuJoCo, then run it for real, with plain conversation as the only control.
Why Practice In A Virtual World
Robots learn slowly and break things when they get it wrong. MuJoCo, a free physics simulator that DeepMind maintains, lets a robot try a move thousands of times on screen before it touches anything real. Pairing that with a chat model means a person can describe a task instead of writing robot code line by line.
Speed matters here. DeepMind described Gemini 4 Hydrogen on Nov. 12 as built for rapid coding and everyday workflows, finishing tasks 2.5 times faster than Gemini 3.8 Flash. A faster model means less waiting between "pick up the cup" and the arm actually moving.
| Date | Google DeepMind Physical AI Step |
|---|---|
| Nov. 10 | Gemini Robotics 2.5 learns hundreds of everyday skills, from folding laundry to sorting car parts |
| Nov. 12 | Gemini 4 Hydrogen launches, 2.5 times faster than 3.8 Flash |
| Nov. 16 | Scientists use Gemini 4 Carbon in Antigravity for physics, math and biology problems |
| Nov. 17 | Hydrogen gets a robotics plugin to drive real robot arms by chat |
Also Read: Humanoid Robots Running Google’s Gemini Just Did Factory Work At Mercedes-Benz, And Got 96% Right
Google's Bigger Robot Push
The plugin is one piece of a fast-moving week. A week earlier, DeepMind showed Gemini Robotics 2.5 learning hundreds of everyday skills, from folding laundry and serving pastries to sorting car parts on a factory floor. Antigravity, first pitched as an AI coding tool, is turning into a single workspace where Google's models write code, run science experiments and now drive hardware.
What To Watch Next
The open question is which robots work with it. DeepMind has not named the supported arms or given a timetable for a wider rollout, and those details will decide whether this stays a lab toy or reaches factories and small workshops.
The bigger race is over who becomes the default brain for robots. If anyone with Antigravity can teach an arm by chatting, Google could win robot builders the same way it wants to win coders, by making its models the easiest place to start.
Read Next: RoboScript Teaches Robots A Task From One Demo, Beating Trained Rivals By 20 Points
Originally published on Fathom News.

