# Google releases Gemini 4 Argon for robot control and policy improvement

published: 2026-11-09T16:05:52.219Z
topic: infrastructure
source: https://www.infoq.com/news/2026/11/google-gemini-argon/?utm_campaign=infoq_content&utm_source=infoq&utm_medium=feed&utm_term=AI%2C+ML+%26+Data+Engineering
url: https://aiagentnews.io/stories/google-releases-gemini-4-argon-for-robot-control-and-policy-improvement

Google has released Gemini 4 Argon, a model it positions for robot control, simulation scene building, robot video labeling and iterative policy improvement. It is available to Google AI Ultra subscribers and through the paid Gemini API under the model id gemini-4-argon.

Google presents Gemini 4 Argon as a general-purpose physical-reasoning model rather than a single end-effector policy. Its published prompts cover robot control, building simulation scenes from teleoperation video, labeling robot footage for behavior cloning, and iterative policy improvement. In one example, Argon directs a Boston Dynamics Spot robot through a foot-pedal task after practice in a MuJoCo twin.

Google says it synthesized domain-randomized MuJoCo episodes from one RGB photo to fine-tune a vision-language-action policy, raising success on the matching real setup from 56% to 75%. It also reports 87% on the ERQA embodied reasoning benchmark using pure reasoning without helper tools, and says that leads other frontier models on that test. The robotics prompts depend on long task-specific instructions, coordinate grids, motion primitives and verifiers, and Google does not publish per-demo failure rates beyond that single VLA case.

Argon is available to Google AI Ultra subscribers and through the paid Gemini API, with the model id gemini-4-argon listed in the Gemini API models documentation. Early reaction from developers focused on simulation fidelity, with one commenter arguing that contact events need to be matched in simulation twins, not just visuals.

## Why it matters if you build agents

Teams building robot agents can now test one hosted model for control, scene construction and labeling instead of stitching together separate VLA, simulator and labeling services. The reported results rely on heavy prompt scaffolding and company-reported benchmarks, so checking simulated contact behavior against real hardware is worth doing before building a pipeline on it.