Google has published a detailed case for using Gemini 4 Argon, the frontier model it launched on 2 October, as the reasoning layer in robotics. The post, written by Google DeepMind principal engineer Kanishka Rao, shows the model controlling a Boston Dynamics Spot quadruped inside the MuJoCo simulator, steering a Franka FR3 arm with no task-specific training, and building calibrated simulation twins from robot camera feeds, human video and hand-drawn sketches.
Google reports Argon scoring 87% on ERQA, its Embodied Reasoning benchmark, which it says beats every other frontier model. Two self-improvement results are more concrete. Generating over 4,500 simulated episodes lifted a real-world task success rate from 56% to 75% and cut completion time from 35 seconds to 24. Three rounds of iteration on an SO-101 arm doing pick-and-place cut completion from 58.9 seconds to 17.4.
Why it matters: the figures are Google’s own and unaudited, and ERQA is a benchmark Google both built and now tops. The practical claim is the more interesting one, that a general model can manufacture its own training data in simulation instead of needing more teleoperated demonstrations. Argon is available through the Gemini API on the paid tier and in AI Studio for Google AI Ultra subscribers.