Google Releases Gemini 4 Argon for Robot Control and Simulation
Google released Gemini 4 Argon, positioned as a general-purpose physical reasoning model for robot control, Real2Sim scene construction, and iterative policy optimization rather than a single end-effector policy. Google reports that after training in a MuJoCo twin environment, Argon guided Boston Dynamics' Spot through a foot-pedal task, and that domain-randomized MuJoCo scenes synthesized from a single RGB photo lifted a vision-language-action policy's real-world success rate from 56% to 75%; it also scored 87% on the ERQA embodied reasoning benchmark using pure reasoning. The model is available to Google AI Ultra subscribers and via the paid Gemini API, though the demos rely on lengthy task-specific prompts and verifiers, and failure rates were not disclosed.
Full Report
Google has released Gemini 4 Argon, positioning it as a general-purpose physical reasoning model for robot control, Real2Sim scene construction, robot video multimodal annotation, and iterative policy optimization, rather than a single end-effector policy.
In the published prompts, after training in a MuJoCo twin environment, Argon guided Boston Dynamics' Spot robot through a foot-pedal task; used visual coordinates and HTTP motion primitives to operate a relatively simple robotic arm; and achieved closed-loop control via the Gemini Omni video world model. The Real2Sim example reconstructed a MuJoCo scene from teleoperation video; the annotation prompts annotated demonstration videos to enable behavior cloning. Google also reported that it synthesized a domain-randomized MuJoCo scene from a single RGB photo, using it to fine-tune a vision-language-action policy, and noted that in the corresponding real environment, the success rate improved from 56% to 75%.
Google reported that on the "Embodied Reasoning Question Answering (ERQA)" benchmark, Argon scored 87% using pure reasoning alone (without auxiliary tools), and stated that this result leads other frontier models on this test. Argon is available to Google AI Ultra subscribers and can also be accessed through the paid Gemini API, with the model ID gemini-4-argon (see the Gemini API model documentation). The robotics blog post uses lengthy task-specific prompts, coordinate grids, motion primitives, and verifiers, so the auxiliary structure becomes part of the system under test, and Google only published a single VLA case without publishing failure rates for the various demonstrations.
Early practitioners' reactions focused mainly on simulation fidelity. Davin YC Dong commented on X:
Google's Gemini 4 Argon robotics demo cases not only calibrate the simulation twin from a visual perspective, but also specifically tune for real-world physical contact events. This is exactly the part I would focus on checking: a scene may look realistic, but the physics behavior within it may be incorrect. For robotic manipulation, matching the actual behavior at contact is just as important as matching pixels.
In addition, regarding the Video-to-Sim-to-Real claim circulating this week, a developer named pulpmatrix called for rigor regarding sources:
Video-to-Sim-to-Real technology is real, and Google's developer blog Gemini 4 Argon for Robotics has confirmed this (MuJoCo twin models, ERQA reaching 87%). However, the Cornhole game is not included: there is neither a paper nor a codebase, nor hit-rate data, and the actual throwing data was first used to calibrate the robot.
Dedicated vision-language-action systems, such as OpenVLA-style policies, Physical Intelligence's π series models, and NVIDIA GR00T with its accompanying Isaac toolchain, are optimized for robot action prediction in embodied scenarios, typically relying on large teleoperation or simulation corpora. Scene-to-sim and generative world tools, World Labs-style reconstruction, NVIDIA DreamDojo-style action videos, and Isaac Sim digital twins focus on environment synthesis or video prediction, but do not necessarily have the complete control and quality control (QC) loop demonstrated in Google's prompts. In the field of spatial perception and planning, Google's previous work, such as Gemini Robotics-ER, has consistently been an important line of research. From an architecture review perspective, the key question is: can a frontier model plus a complex prompt system replace a pipeline composed of VLAs, simulators, and annotation services?
Disclaimer: This article is an InfoQ translation, and reproduction without permission is prohibited.
Original link: https://www.infoq.com/news/2026/11/google-gemini-argon/