As reported on November 20, 2026, Google had introduced Gemini 4 Argon as a general-purpose physical reasoning model for robotics. Its stated uses span robot control, reconstruction of simulation scenes from real-world footage, video annotation and iterative policy improvement—not just direct control of a robot’s end effector.
Google’s published examples include guiding a Boston Dynamics Spot robot through a foot-pedal task after training in a MuJoCo simulation twin, controlling a simpler robotic arm through visual coordinates and HTTP motion commands, and using the Gemini Omni video world model for closed-loop control. Another example reconstructs a MuJoCo scene from teleoperation video; an annotation workflow labels demonstration footage for behavior cloning.
Google also reported fine-tuning a vision-language-action policy with domain-randomized MuJoCo scenes synthesized from a single RGB photo. It said the policy’s success rate in the corresponding real-world setting rose from 56% to 75%. On the Embodied Reasoning Question Answering (ERQA) benchmark, Google reported an 87% score using reasoning alone, without auxiliary tools, and said the result led other frontier models on that test. These are Google-reported results, not independently established performance across robotics tasks.
The demonstrations rely on substantial task-specific prompts, coordinate grids, motion commands and verification steps. Those components are part of the systems being demonstrated, making it difficult to attribute outcomes to the model alone. Google published a single vision-language-action policy case and did not provide failure rates across the demonstrations.
Practitioners raised questions about what the examples establish. Davin YC Dong argued that a simulation twin must match real-world contact behavior, not merely look accurate. A developer posting as pulpmatrix cautioned against conflating Google’s documented video-to-simulation-to-real workflow with circulating claims about a cornhole demonstration, for which they said there was no paper, repository or hit-rate data.
Gemini 4 Argon was available to Google AI Ultra subscribers and through the paid Gemini API under the model ID gemini-4-argon. The open question for robotics teams is whether a general-purpose model paired with extensive prompts can replace parts of a pipeline built from specialized action models, simulators and annotation tools; the published examples do not settle that question.



