Google's Gemini 4 Argon brings deep physical reasoning to real robots

Google has unveiled a robotics-focused version of its latest AI model, Gemini 4 Argon, designed to help machines understand and interact with the physical world.

Announced October 13, 2026, on the Google Developers Blog, Argon is described as a general-purpose physical reasoning model that covers the entire robotics lifecycle — from planning and simulation to real hardware control and self-improvement.

The model was tested on the Embodied Reasoning (ERQA) benchmark, which measures how well an AI can solve physical and spatial problems in 3D environments. Using pure reasoning alone, Argon scored 87%, outperforming every other frontier model ¹.

Real robots, real tasks

Google demonstrated Argon controlling a Boston Dynamics Spot quadruped robot — a four-legged machine with 12 degrees of freedom in its legs and a 6-degree-of-freedom arm ¹.

In one demo, Spot was asked to pick up soda cans and throw them in a foot-pedal trash bin. Rather than acting blindly, Argon first created a digital twin of the trash bin inside MuJoCo, a physics simulation engine, and practiced the task virtually before executing it on the real robot ¹.

The system also handled zero-shot control — meaning it completed a task it had never been explicitly trained for — by uprighting a fallen mug and placing a crumpled napkin inside it, using a calibrated wrist camera with a 2 cm coordinate grid ¹.

Why it matters

Most AI models are trained to understand text or images. Argon is built to reason about physical objects in three dimensions — their positions, movements, and how they interact. This is a critical step toward physical artificial general intelligence, or physical AGI, where machines can operate reliably in the real world, not just in digital environments.

Google outlined four areas where Argon can be applied: robot control, building simulation twins from real camera footage or sketches, understanding physical dexterity and progress, and improving actions through simulation and self-learning on actual hardware ¹.

Available now

Gemini 4 Argon is now available for developers and enterprises through the API on Google AI Studio and the Gemini platform. It is also rolling out to subscribers of Google's paid AI tiers ⁴.

The broader Gemini 4 series was introduced earlier as a frontier model for coding, enterprise knowledge work, and cybersecurity ⁵. The Argon variant adds specialized capabilities in robotics and physical reasoning on top of those general skills.

Google says the model can reconstruct physically accurate simulation environments from robot camera feeds, human videos, or even hand-drawn paper sketches, opening new pathways for training robots without expensive real-world setup ¹.

Key facts

• Gemini 4 Argon scored 87% on the ERQA 3D physical reasoning benchmark
• It controls Boston Dynamics Spot robots for real-world manipulation tasks
• Argon creates digital simulation twins to practice skills before using real hardware
• The model is available now via Google AI Studio and the Gemini platform
• Four application areas: robot control, Real2Sim, physical understanding, and self-improvement

Sources

1. Gemini 4 Argon for Robotics- Google Developers Blog
2. Gemini 4 Argon for Robotics - Google Developers Blog
3. Gemini 4 Argon - Model Card — Google DeepMind
4. Gemini 4 Argon is now available for everyone - The Keyword
5. Introducing Gemini 4 Argon - The Keyword