Google pitches Gemini 4 Argon as a unified AI model for robotics pipeline

Google is pitching Gemini 4 Argon as a single, general-purpose AI model to manage the full robotics pipeline, controlling robots, building virtual practice worlds, and grading training videos. For instance, when controlling a Boston Dynamics Spot robot to toss soda cans, Argon first builds a virtual copy of the bin and uses simulation practice before real-world execution, with two extra AI agents verifying positioning. From a single photo of party props, Argon generated over 4,500 training episodes, boosting the real robot's success rate from 56% to 75% and reducing execution time from 35 to 24 seconds. Argon can also generate simulations from teleoperated recordings, first-person videos, or even ballpoint pen sketches, aiming toward 'physical AGI'.