Google unveils Gemini 4 Argon for Robotics

Google unveiled Gemini 4 Argon for Robotics, a model for understanding and controlling robots' physical situations, building simulations and improving motions, on October 13 local time, according to aitimes.com.

The model reasons over video, 3D space and physical laws together to help robots perform complex tasks and to practice motions in simulation. It scored 87% on the Embodied Reasoning Question Answering (ERQA) benchmark using reasoning alone without separate tools.

Unlike Google DeepMind's RT series (RT-1, RT-2) VLA models, Argon is described as a higher-level reasoning and simulation design engine rather than a terminal model that directly controls a robot.

It automatically builds physics simulation environments such as MuJoCo and generates large training data to maximize the task success rate of general-purpose VLA models. Generated simulation training data raised a general-purpose VLA model's task success rate from 56% to 75%.

In a demonstration, it controlled Boston Dynamics' quadruped robot Spot to step on a pedal trash can, open the lid and throw away a drink can. The model is offered to AI Ultra subscribers in Google AI Studio and Google Antigravity, with related developer tools and SDKs for robot control and simulation applications.

Background

Google DeepMind previously developed the RT series (RT-1, RT-2) as standards for robot control, VLA (Vision-Language-Action) models that learn from large-scale robot navigation and manipulation demonstration data and issue direct driving commands.