Google has released Gemini 4 Argon as a general physical-reasoning model for robotics rather than a single policy tied to one robot. Its examples span robot control, rebuilding simulated scenes from video, labeling demonstrations and using failures to improve a vision-language-action policy.
Published workflows show Argon directing a Boston Dynamics Spot after practice in a MuJoCo digital twin and controlling a simpler arm through visual coordinates and motion commands. Google also generated varied simulated episodes from one RGB image to fine-tune a policy, reporting that success on the corresponding real setup rose from 56% to 75%.
Google says Argon scored 87% on the Embodied Reasoning Question Answering benchmark without helper tools. The model is available to Google AI Ultra subscribers and through the paid Gemini API under `gemini-4-argon`.
The demonstrations rely on long task-specific prompts, coordinate grids, verifiers and motion primitives. That scaffolding is part of the system, so the results should not be attributed to the base model alone. Google also did not publish failure rates across every demonstration beyond the single policy-improvement case. Engineering teams will need to test whether one frontier model and a substantial prompt harness is more reliable than separate control, simulation and labeling components for their robots.