---
title: "Google Gemini 4 Argon: Flagship AI Model That Teaches Robots to Think and Act Without Prior Training"
description: "Google introduced Gemini 4 Argon, a universal AI model for robotics that scores 87% on the ERQA benchmark and is capable of controlling robots in zero-shot mode without prior training."
date: 2026-10-17T02:32:26.913Z
lang: en
url: https://xab.info/en/posts/google-gemini-4-argon-robotics-embodied-ai
tags: [google, gemini-4-argon, robotics, embodied-ai, artificial-intelligence, boston-dynamics, mujoco, zero-shot]
author: "Alexander Buran"
publisher: "XAB.info"
---

# Google Gemini 4 Argon: Flagship AI Model That Teaches Robots to Think and Act Without Prior Training

![Abstract blue image featuring a minimalist robotic arm icon](https://xab.info/media/2026/10/17/google-gemini-4-argon-robototekhnika-embodied-ai/google-gemini-4-argon-robototekhnika-embodied-ai-1.webp)

## Key Points

- Gemini 4 Argon scored 87% on the Embodied Reasoning (ERQA) benchmark, outperforming all other advanced models
- The model enables robots to execute tasks in zero-shot control mode without prior training
- Google demonstrates control of Boston Dynamics' Spot robot and a robotic arm via Argon
- Real2Sim technology converts real scenes and even hand-drawn sketches into physical 3D simulations
- Argon is positioned as a universal model of physical reasoning rather than a replacement for specialized Gemini Robotics 1.5

In September 2026, Google introduced its flagship artificial intelligence model, Gemini 4 Argon, which the company states is designed not only for human-computer interaction but also for controlling robots. In its corporate blog, Google described Argon as a "qualitative leap in robotics" capable of accelerating progress toward general-purpose physical, or embodied, artificial intelligence. The key differentiator of the model is that robots equipped with Argon are able to think, learn, and adapt to new tasks without significant fine-tuning costs—meaning they can execute actions in a zero-shot control mode.

### Embodied Reasoning Benchmark Results

Google substantiated its claim of Argon's leadership in embodied reasoning with results from the Embodied Reasoning (ERQA) benchmark suite, which tests a model's ability to understand and solve physical and spatial problems in three-dimensional environments. Using pure logical reasoning, Argon scores 87%, outperforming all other advanced models on the market. According to the developers, this metric demonstrates that the model is capable of sequential physical reasoning without the need for training on task-specific datasets.

### Relation to Existing Google Robotics Models

It is important to note that at the time of publication, Google has not yet tied the adaptation of Gemini 4 Argon or its components to physical embodiment AI platforms. The company separately develops two specialized robotics models—Gemini Robotics 1.5 and Gemini Robotics-ER 1.5—which jointly implement a "think first, act later" approach where the robot first generates a plan of action and then executes it. Argon does something similar, but in a more universal format: the model combines an understanding of physical phenomena and processes, code generation, and visual-semantic reasoning into a single architectural solution.

### Practical Examples: From Robot Dog to Real2Sim

In its corporate blog, Google provided illustrative examples of Argon being used in four key areas of robotics: action simulation and real-world embodiment, learning from examples, environment understanding, and skill refinement. The first demonstration case involved controlling Boston Dynamics' Spot robot dog in the MuJoCo physics simulator without prior training. Argon taught the quadruped robot a new skill—pressing a trash can pedal to discard soda cans picked up from the floor. To practice the movement, the AI creates a digital twin of the trash can with a pedal in MuJoCo and only then allows the robot to reproduce the sequence of actions in reality, correcting its movements, tracking the position of the manipulator, the can, and the open can, and helping maintain balance on three limbs while the fourth presses the pedal.

### Zero-Shot Manipulator Control and Real2Sim Technology

In the second experiment, Argon controls a robotic arm to set an overturned mug upright and place a crumpled napkin inside. The system uses spatial coordinates and movement/grasping commands, re-analyzing the image after each step to adjust actions based on object positioning. The manipulator performs all actions without prior training. The third demonstration technology, Real2Sim, converts real-world scenes into physical simulations: Argon analyzes video recordings, reconstructs object geometry, and assigns physical properties (mass, inertia, friction, collision behavior). The virtual scene is then verified against real camera footage. The same process allows a rough hand-drawn sketch—such as a diagram of a robot knocking over a pyramid of blocks with a ball—to be turned into a full-fledged 3D animation with realistic physics.

### Strategic Context: The Path to General AI in the Physical World

According to Google developers, Gemini 4 Argon represents an important milestone on the path toward creating general artificial intelligence capable of operating in the physical world. The new model integrates programming, visual perception, physics modeling, and real robot control into a unified architecture. By developing Argon, the company states its intention to build robotic solutions that can work alongside humans and help them solve complex everyday and industrial tasks. Thus, Argon is positioned not as a narrowly specialized tool for a specific class of robots, but as a universal model of physical reasoning applicable across all stages of the robotics lifecycle.

## Fact-Check & Sources

- [Google заявила об «умственном» превосходстве роботов с ИИ-моделью Gemini 4 Argon над остальными window-new](https://3dnews.ru/1150007/google-zayavila-ob-umstvennom-prevoshodstve-robotov-s-iimodelyu-gemini-4-argon-nad-ostalnimi) - The primary source contains all key facts: ERQA results (87%), Spot and manipulator examples, Real2Sim description, and mention of Gemini Robotics 1.5. All data has been transferred to the article without distortion.

## FAQ

### What is Gemini 4 Argon and when was it introduced?

Gemini 4 Argon is Google's flagship AI model introduced in September 2026. It is designed for both human interaction and robot control, combining physics modeling, code generation, and visual-semantic reasoning.

### What results did Argon show in the ERQA benchmark?

Using pure logical reasoning, Gemini 4 Argon scores 87% in the Embodied Reasoning (ERQA) suite, outperforming all other advanced models on the market.

### How does Argon differ from Gemini Robotics 1.5?

Gemini Robotics 1.5 and Gemini Robotics-ER 1.5 are specialized robotics models implementing a "think first, act later" approach. Argon is more universal: it encompasses physical understanding, code generation, and visual-semantic reasoning, and can be applied across all stages of robotic system lifecycles.

### What is zero-shot control in the context of Argon?

Zero-shot control is the model's ability to control a robot without prior training on specific tasks. In Google's demonstrations, Argon controlled the Spot robot and a robotic arm, performing novel actions (pressing a pedal, setting a mug upright) without a fine-tuning stage.

### What is Real2Sim technology?

Real2Sim is a technology where Argon analyzes video recordings of real scenes, reconstructs object geometry, and assigns physical properties (mass, inertia, friction). A virtual simulation is then created and verified against real images. The technology also allows hand-drawn sketches to be turned into 3D animations with realistic physics.