Google is positioning Gemini 4 Argon as a single model for several stages of robotics work: planning actions, building simulations for practice, and reviewing training videos. The demonstrations range from a robot learning to handle a trash bin to simulated recreations of published experiments.
The breadth is striking, but the examples also show how much setup people provide. Prompts can include prepared commands, camera images and rules that constrain the model's choices. The gap between an appealing demonstration and reliable robot performance remains central to judging the system.
From a photo or sketch to a practice environment
In one demonstration, a Boston Dynamics Spot robot is tasked with putting soda cans into a step-on trash bin. Argon creates a virtual version of the bin so the robot can practice balancing on three legs before trying the task on real hardware. Two additional AI agents check the robot's foot and gripper positions before it drops the can.
Other examples start with a tabletop photo, a recording of a robot controlled remotely, or a first-person video of a person grasping an object. Argon uses these inputs to build simulations, reproduce aspects such as fingertip force readings, or map human hand and finger movements onto a virtual humanoid robot.
A hand-drawn sketch can also serve as a starting point. In one case, Argon turns a drawing of a humanoid throwing a ball at boxes into a simulation that works out a motion to knock down all six boxes while keeping the robot upright. These examples show the model translating different kinds of visual material into spaces where actions can be explored.
For a table scene with party props, Google says Argon generated 77 tasks and more than 4,500 training episodes from a single photo. On the real robot, the reported success rate rose from 56 to 75 percent, while execution time fell from 35 to 24 seconds. The source does not state how many attempts underlie that success rate.
Planning actions and learning from retries
Argon can also direct Gemini Omni, a video world model that generates clips showing how a scene changes after an action. Argon issues timestamped instructions, checks frames from the resulting video and adjusts its next instruction. In one demonstration, the task was to arrange five scattered letter tiles to spell “ARGON”; the sequence took 33 steps.
A separate robot-arm example shows a more iterative process. The task was to place a Lego brick in a wooden box, with instructions for Argon to inspect logs and camera images after a failure, identify a cause and retry with a verified fix. The first run took just under a minute. The second took 23 seconds but clipped the box edge; on the third attempt, the brick landed in the center after about 17 seconds.
Google also says Argon can turn a prompt based on a research paper into an interactive browser simulation. The company reports that it has built 68 such sandboxes, with examples including robot soccer and a humanoid on a skateboard. The goal is to make robotics experiments easier to explore through simulations.
Benchmarks tell different stories
Google reports that Argon scored 87 percent on the ERQA benchmark, which tests reasoning about physical and spatial problems in 3D environments. The test used 400 questions and no helper tools. The reported comparison scores were 82.3 percent for Claude Opus 5.5, 81 percent for GPT-6 Astra and 74 percent for Claude Fable 5.1.
A separate evaluation offers a less favorable result. In Artificial Analysis's AA-Robotics benchmark, language models controlled a real robot arm across eight tasks. Argon scored 55 out of 100 and placed third, behind GPT-6 Astra at 69 and GPT-6.1 Sol at 65. The contrast matters: strong results on spatial reasoning questions do not by themselves establish the same strength in direct robot control.
Data review, prompts and practical limits
Argon is also intended to assess the quality of training videos. In a video just over two minutes long, it divides a dual-arm robot's origami fox folding into sub-steps and flags pauses and poor grips, which are then given less weight during training. In another example, it tracks where a person sorting rings makes progress, stalls or reverses direction.
The examples depend on carefully described tasks. For a cup-righting exercise presented as “zero-shot,” Argon receives a camera image with a centimeter grid and prepared commands such as “grasp” and “lower until contact.” Simulation prompts also contain long lists of restrictions designed to prevent shortcuts, including copying measured sensor values.
Google's blog post does not give failure rates for individual demonstrations, and the reported 75 percent success rate comes without the total number of attempts. Those omissions make it difficult to assess how consistently the demos would transfer to other tasks.
Argon is available through Google AI Studio and Antigravity for Google AI Ultra subscribers, and through the paid tier of the Gemini API. Google's distinguishing claim is that one model can span simulation, control and evaluation. The demonstrations make that ambition concrete, while the prompt details and mixed benchmark results leave reliability as an open question.