Google has turned its Antigravity coding agent into a robotics workbench. A new Robotics plugin, announced on 17 November, lets the agent set up a real robot arm, test it in simulation and run it on hardware, all from plain-language prompts inside the coding environment.
Hydrogen, built for speed
At the centre is Gemini 4 Hydrogen, Google’s fast model for spatial and embodied reasoning. It carries the skills of the company’s Gemini Robotics ER 2 models into the Gemini 4 line, and Google says it outperforms that older model on the public ERQA benchmark for spatial reasoning.
On the Artificial Analysis Robotics 0.1 benchmark, Hydrogen steers a real six-axis arm through eight everyday tabletop tasks, from pouring a drink to packing blocks into a box. It scores 58, similar to the heavier Gemini 4 Argon, but at a lower cost per token.
A plugin that runs the robot
The Robotics plugin packages the surrounding work into skills Gemini can call on the developer’s behalf. It covers robot bring-up, MuJoCo simulation, perception and control, and application building, turning scattered drivers, calibration scripts and simulators into one guided workflow inside Antigravity.
Starting from an empty desk, a developer plugs in an I2RT YAM arm and an external depth camera, then asks Hydrogen to bring it up. The plugin installs the software, finds the camera, calibrates the arm and runs test motions, leaving it ready for tasks in about ten minutes.
The same flow also brings up a low-cost SO-ARM101 arm and a Franka FR3 Duo, from scratch to live waving. Support for several arm types matters, because robotics work has long stalled on the hours lost to drivers, wiring and calibration before any real task begins.
Simulation before hardware
Once the robot is live, MuJoCo runs directly inside Antigravity, with skills for creating, auditing and rendering simulated scenes. Developers can describe a workspace, import it from a document or photo, and adjust shapes and physics through chat before deploying anything to physical hardware.
Two working examples
Google’s worked example starts with a whiteboard sketch of a conveyor carrying mixed snacks. From one prompt, Hydrogen builds a MuJoCo simulation, tests the grasping policy across multiple runs, then deploys the verified policy to the real arm, which sorts snacks off the moving belt within about an hour.
A second example builds a small web storefront for a soap shop, then connects two YAM Pro arms that pack orders when they arrive. A third has the FR3 Duo simulate a chess board and play against itself or a human opponent, showing the same loop on very different hardware.
Availability
To use it, developers select gemini-4-hydrogen in Antigravity and install the Robotics plugin from the Build With Google shelf. Antigravity itself needs a paid Google AI Pro or Ultra subscription, while the model is also available through the Gemini API and Google AI Studio.
Our Read
This is Google moving its coding agent from software into the physical world. Antigravity already writes and tests code; now it calibrates hardware, runs simulations and closes the loop on a real arm. That turns a developer tool into a robotics bench for everyday use.
Why It Matters
Robotics has been held back less by model intelligence than by integration work. If an agent can handle drivers, calibration, simulation and control from one prompt, small teams can prototype physical products the way they prototype software, and Google’s lead in embodied models becomes an advantage.
Background
Antigravity has expanded quickly since Google folded its Jules coding agent into the platform and opened a plugin marketplace in October. Voice mode and remote agents followed in November, then a layout for knowledge work. Robotics is the platform’s next step outward.
Source: Google Developers Blog


0 Comments