Google DeepMind has developed a robot plug-in for its own AI programming tool Antigravity. The underlying model is Gemini 4 Hydrogen. What it wants to solve is the most annoying preparation work in robot development.
The process is as follows: plug in a supported robotic arm and a camera, and the AI will install the driver, find the camera, and calibrate the robotic arm. This step takes about 10 minutes in the Google demo.
Then you describe in plain English what you want it to do. The model will be set up in the MuJoCo physics simulator to set up a practice scene, write control code, and then continue to adjust until the task is completed.
An example given by Google: It took a developer about an hour from sketching on a whiteboard to letting a robotic arm sort snacks from a moving conveyor belt.
Currently supported robotic arms include entry-level models such as the very cheap SO-ARM101.
The threshold for this matter has been lowered significantly. In the past, most people who played with robotic arms could get stuck just by driving and calibrating them, let alone writing motion control and adjusting simulation parameters. Now this layer is eaten away by AI, and developers only need to figure out the task itself.
The same model was also used to do something interesting: using a Google Sheet as a canvas, shrinking thousands of cells into pixels, and making animated pixel paintings, including a train entering the station, a bicycle riding past a chapel, and autumn leaves falling. The file name said "100% formula."
Looking at the two things together, Hydrogen is positioned by Google as an execution model and is good at breaking down a sentence into specific operations, whether the operation object is a form or a robotic arm.