Gemini 4 Argon for Robotics Blog (developers.googleblog.com)

Gemini 4 Argon has been unveiled as a transformative tool for robotics, enhancing physical understanding, code generation, and visual-semantic reasoning, driving closer to achieving physical AGI. It excels in the Embodied Reasoning (ERQA) benchmark, scoring an impressive 87%—the highest among current models—demonstrating its advanced capabilities in solving physical, spatial challenges in 3D environments. This model facilitates a more dynamic approach across the robotics lifecycle, enabling tasks such as teaching a Boston Dynamics Spot® robot to manipulate objects with precision and develop interactive digital twins for real-world robotics applications. Argon’s approach allows it to apply reasoning in real-time and adapt to complex tasks through a closed-loop system, refining its prompts based on iterative visual feedback. It leverages detailed simulations in MuJoCo, synthesizing accurate models that closely mirror real-world dynamics, and generates refined robotic trajectories for various manipulations. These capabilities position Argon at the forefront of robotics, promising a leap in automating complex manipulation tasks and enhancing robotic dexterity and efficiency in real-world scenarios. Its ability to analyze and improve robotic actions through visual reasoning offers significant implications for future developments in autonomous systems and robotics.