Google Unveils Gemini 4 Argon: A Unified Model for Robot Simulation, Control, and Evaluation

With Gemini 4 Argon, Google is pitching a single model for the full robotics pipeline. It controls robots, builds virtual practice worlds, and grades training videos. But the example prompts reveal just how much human hand-holding the model still needs.

A Boston Dynamics Spot robot is supposed to toss soda cans into a step-on trash bin. Before it touches the real pedal, Gemini 4 Argon builds a virtual copy of the bin, and the robot practices balancing on three legs inside that simulation. Only then does it attempt the task on real hardware. Before it drops the can, two extra AI agents verify that the foot and gripper are positioned correctly.

That's how Google demos the abilities of its newest model for robotics tasks. Argon is designed to better grasp spatial relationships, write control code, and interpret videos in terms of physics. Google sees the model as a general-purpose tool that covers everything from controlling individual robots to recreating published research, and calls it another step toward "physical AGI," a goal it and its AI subsidiary DeepMind have been pursuing for years.

Gemini 4 Argon controls Spot to pick up cans and toss them into a trash bin using the pedal, after practicing the sequence in a MuJoCo simulation. | Image: Google

A single photo turns into thousands of training runs

The Spot demo follows a pattern that shows up across many of Google's examples: Argon uses simulations to prepare tasks, verify them, or generate training data. From a single photo of a table with party props, the model built a virtual replica, including a soft plush cap. It came up with 77 tasks, from righting toppled objects to clearing the table, and generated more than 4,500 training episodes for a control model. When tested on the real robot, the success rate climbed from 56 to 75 percent and execution time dropped from 35 to 24 seconds.

From a single photo, Argon generates over 4,500 varied simulation episodes that boost the real robot's success rate from 56 to 75 percent. | Image: Google

Argon can work from other sources, too. From a recording of a teleoperated robot, it builds a simulation designed to look the same and replicate the same force readings at the fingertips. From a first-person video of a human performing a grasping task, the model maps hand and finger movements onto a virtual humanoid with five-fingered hands.

Even a ballpoint pen sketch works. One drawing shows a humanoid robot knocking down a pyramid of boxes with a ball, and Argon turns it into a simulation that calculates a throwing motion to topple all six boxes without the robot falling over.

From a simple hand sketch, Argon builds a simulation where a humanoid robot knocks down all six boxes with a single throw while keeping its balance. | Image: Google

Video 5

Argon directs an AI video model to plan robot actions

One demo plays out entirely inside an AI-generated video. Gemini Omni is a video world model that generates clips showing how a scene changes when something happens in it, and Argon acts as the director. It writes timestamped instructions describing what the hands should do, like sliding a Scrabble tile to the center of the table. Gemini Omni generates a video from those instructions, and Argon then checks individual frames and adjusts the next command. The goal is to arrange five scattered letter tiles to spell "ARGON," and after 33 steps, they're in the right order.

Argon directs the video world model Gemini Omni step by step until the robot hands have sorted the letters to spell ARGON. | Image: Google

In a separate demo with a simple robot arm, Argon had to place a Lego brick into a wooden box and improve its own performance over time. When attempts failed, the prompt told Argon to review its own logs and camera images, find the cause, and retry after applying a verified fix. The first run took just under a minute, and the second was much faster at 23 seconds, though the brick clipped the edge of the box. On the third try, it landed dead center after about 17 seconds.

Over three iterations, Argon cuts the robot arm's execution time from 58.9 to 17.4 seconds and hits the center of the box on the final try. | Image: Google

Google also wants Argon to serve research. Given a single prompt per paper, it recreates published robotics experiments as interactive simulations that run in the browser. Google says 68 such sandboxes have been built so far, including robot soccer and a humanoid robot on a skateboard.

Video 6 The ERQA benchmark (Embodied Reasoning Question Answering) tests how well AI models understand and solve physical and spatial problems in 3D environments. The test used 400 questions with no helper tools, so models had to solve everything through reasoning alone. Argon scores 87 percent according to Google, ahead of all other frontier models. The runner-up, Claude Opus 5.5, hit 82.3 percent, about five percentage points behind. GPT-6 Astra followed at 81 percent, and Claude Fable 5.1 trailed at 74 percent, 13 points back.

On the ERQA benchmark for spatial and physical reasoning, Gemini 4 Argon leads Claude Opus 5.5 by about five percentage points, according to Google. | Image: Google

The AA-Robotics benchmark from Artificial Analysis, published a day later, tells a different story. In that test, language models directly control a real robot arm across eight tasks, and Argon landed third with 55 out of 100 points, behind GPT-6 Astra at 69 and GPT-6.1 Sol at 65.

Argon also reviews training data quality

Beyond controlling robots and building simulations, Argon is meant to evaluate training data. In a video just over two minutes long showing a dual-arm robot folding an origami fox, the model breaks the sequence into sub-steps and flags pauses and botched grips. Those segments are then weighted less during training. In another video of a human sorting rings, Argon traced where the task progresses, stalls, or goes backward.

Google also showed a more playful use case built around a classic weighing puzzle. Among twelve identical-looking balls, one is heavier or lighter than the rest. Argon built a simulation with two robot arms and a balance scale, reading the scale's tilt from camera images alone. It finds the odd ball in three weighings or fewer.

The real scaffolding lives in the prompts

Google published detailed example prompts for the demos, and they put some of the bigger claims into perspective. For righting an overturned cup, described as "zero-shot," Argon receives a camera image with a centimeter grid overlay and pre-built commands like "grasp" or "lower until contact." For the simulations, there are long prohibition lists to keep the model from taking shortcuts, like simply copying measured sensor values. Google doesn't share failure rates for individual demos in the blog post, and the 75 percent success rate lacks any mention of how many total attempts it's based on.

Without prior training, Argon rights a toppled cup and tucks a napkin inside, guided by a coordinate grid in the camera image. | Image: Google

Argon is available now in Google AI Studio and the development environment Antigravity, but only for Google AI Ultra subscribers. Developers can also access the model through the paid tier of the Gemini API.

Google had released Gemini 4 Argon globally just days earlier. For robotics, the company had already introduced Gemini Robotics-ER 1.6, a model for spatial perception and task planning. Training in virtual environments isn't new either: the startup World Labs showed how real scenes can be turned into virtual training worlds, while Nvidia's DreamDojo simulates robot actions as videos. What sets Google's approach apart is the ambition to handle simulation, control, and evaluation all with the same model.