Gemini 4 Hydrogen drives a real robot arm by conversation
Google DeepMind's new Antigravity plugin connects a supported arm, rehearses in MuJoCo, then runs the task for real.
Google DeepMind has released a robotics plugin for Gemini 4 Hydrogen that lets you control a real robot arm by talking to it. The pitch from the company is short. Connect a supported arm, practice in MuJoCo simulation, and run tasks in the real world using plain conversation. The plugin runs inside Google's Antigravity.
What the plugin does
The loop is the interesting part. Rather than fine-tuning a policy on your hardware first, you plug the arm in, prompt the model, and let it rehearse the task in simulation before it touches anything physical. Google DeepMind frames this as teaching robots new tasks getting a lot faster.
Keerthana Gopalakrishnan, posting as @keerthanpg, gave the developer version of the same claim. She says you can connect your hardware and prompt it to do things in under 10 minutes, with no fine-tuning and no hosting needed. She also says Hydrogen is much faster than Carbon, and aims that at people who found Astra, Carbon and Fable too slow for controlling real robots. The plugin ships with skills for control, simulation and harnesses, and it is on GitHub.
Speed matters here in a way it does not for a chatbot. A model that takes too long to decide what to do next cannot close the loop on a moving arm, which is why latency keeps coming up whenever a general frontier model gets pointed at hardware. That argument has been running all month, including the frontier models versus robot specialists fight at CoRL. This release is the frontier model camp handing out a plug.
Shown versus claimed
The speed claim comes from Google and from a Google researcher. Nobody outside the company has posted their own arm running this yet, so the under 10 minutes number is unverified. The sources do not say which arms are supported, how reliable the resulting behaviour is, or what happens when the sim rehearsal and the real world disagree. They also do not show a task from start to finish.
What is new is the packaging. A general model, a simulator, and a hardware connection in one plugin, aimed at people who want to try something on an arm this afternoon rather than collect a dataset first. Whether that holds up on hardware Google did not choose is the thing to watch, and the GitHub release means somebody will find out quickly.
Vibe robotics is here!
you can just connect your hardware and prompt it to do things in under 10 minutes - no fine-tuning / hosting needed
Sound like you were at the demo
One email a day. Under three minutes. The demos, the deployments and the part the press release left out.
