Source · article
Gemini 4 Argon for Robotics
Summary
Google Developers Blog post presenting Gemini 4 Argon as a general-purpose physical
reasoning model for robotics, usable through the Gemini API (gemini-4-argon, paid
tier, Interactions API). It shows the same general model controlling robots, building
simulations, and generating training data, rather than a separate robotics model.
Key points
- Demos: mobile manipulation on a Spot robot in MuJoCo, zero-shot robotic arm control, steering a video world model.
- Real2Sim: builds simulation twins from robot camera feeds, egocentric video, or sketches.
- ERQA (embodied reasoning): 87% with pure reasoning, which Google says beats every other frontier model.
- Generated 4,500+ randomised sim episodes for a vision-language-action policy; real-world success rose from 56% to 75%.
- Reproduced 68 robotics papers as interactive browser sandboxes from single prompts.
- The post has no robotics-specific safety section; safety appears only as constraints in example prompts.
Why it matters
Frontier general models are now pitched directly for robotics work, competing with dedicated robot foundation models.