Google has announced that its Gemini 4 Argon AI model, introduced in September, is designed to enhance robotic capabilities. The model allows robots to learn, adapt, and perform tasks without extensive prior training.
In tests using the Embodied Reasoning (ERQA) benchmark, which evaluates spatial and physical problem-solving in 3D environments, Gemini 4 Argon achieved an 87% score, outperforming other advanced models.
The model integrates physical reasoning, code generation, and visual-semantic analysis. It functions differently from Google's existing Gemini Robotics 1.5 and Gemini Robotics-ER 1.5 platforms.
Practical applications include controlling a Boston Dynamics Spot robot in the MuJoCo simulation environment to perform tasks like disposing of trash. The system creates digital twins to refine movements.
Additional tests involved a robotic arm manipulating objects like cups and napkins using zero-shot control. Furthermore, the Real2Sim technology allows the model to convert real-world video into physical simulations.
