Google ships a robotics plugin for Antigravity and scores Gemini 4 Hydrogen at 58 on a tabletop-arm test
Alex Gruenstein and Gabe Cohen’s post on the developers blog is dated 17 November. The plugin is a set of skills inside Antigravity: bring-up, MuJoCo, perception, and a controller. Hydrogen scores 58 on Artificial Analysis Robotics 0.1, which the post says is similar to Argon at a lower cost per token. Antigravity itself is limited to paid Google AI Pro or Ultra subscribers. The model id in the API is gemini-4-hydrogen.
Short2 minIn-depth
Priya Shah2 min read
- Gemini
- Antigravity
- Robotics
- Hydrogen

Photo: GoogleSource page
Google on Tuesday put a robotics plugin inside Antigravity and pointed it at Gemini 4 Hydrogen, the faster model it introduced last week. Alex Gruenstein and Gabe Cohen’s post on the developers blog is dated 17 November. The plugin is a set of skills the model can call: install drivers and calibrate an arm from a USB cable, build and preview MuJoCo scenes, find objects and plan grasps, and generate the controller and the simulation together. The install path they give is Antigravity, then Customizations, then Robotics on the Build With Google shelf. Antigravity, the post says, is for paid Google AI Pro or Ultra subscribers. In the API the model id is gemini-4-hydrogen, with what they call day-one support in the google-genai SDKs.
The number they attach to the model, rather than to the plugin, is on Artificial Analysis Robotics 0.1. Hydrogen controls a real six-axis arm through eight tabletop tasks, from pouring a drink to packing blocks into a box, and scores 58. The post says that is similar to Gemini 4 Argon, at a lower cost per token. It also says Hydrogen brings the skills of Gemini Robotics ER 2 into the main Gemini 4 series and beats those specialized models on nearly every axis they measured. On the public ERQA benchmark, the claim is narrower: Hydrogen outperforms ER 2 on spatial tasks such as multi-view correspondence and reasoning about a 3D trajectory. The recommendation for anyone still calling ER 2 is to try Hydrogen for those workloads. The post does not print ER 2’s score next to the 58.
The demos are timed, and the times are the company’s. A developer plugs in an I2RT YAM arm and a depth camera, and the plugin’s skills install software, find the camera, calibrate the arm, and run test motions in about 10 minutes. The same write-up says a similar flow brings up an SO-ARM101 and an industrial FR3 Duo. A snack-sorting cell, from a whiteboard sketch through a MuJoCo simulation to a belt in the room, is “within about an hour.” A second example connects two YAM Pro arms to a small web shop that packs soap orders. The prompt they suggest for a reader with an SO-101 arm is “Bring up my SO-101 arm and make it wave.” The 58, the model id, and the paid-plan gate are the parts a developer can check without owning an arm. The hour and the ten minutes are the runs Google filmed.


