Google shows Gemini 4 Argon controlling robots and building practice worlds, with reliability unclear

Before a Boston Dynamics Spot robot steps on a trash-bin pedal, Google’s Gemini 4 Argon builds a virtual bin so it can practice balancing on three legs. In Google’s demonstration, the robot then attempts the task on real hardware, with two additional AI agents checking its foot and gripper positions before it drops a can.

That sequence captures Google’s pitch for Argon: one model spanning robot control, virtual training environments and training-data evaluation. Its robotics demonstrations show those jobs feeding into one another, rather than presenting the model only as a system that tells a robot where to move.

Reasoning questions and robot arms give different answers

Google reports that Argon scored 87% on ERQA, a 400-question test of physical and spatial reasoning completed without helper tools. That put it ahead of Claude Opus 5.5 at 82.3% and GPT-6 Astra at 81% in Google’s comparison.

The Decoder reports a different ranking on Artificial Analysis’s AA-Robotics benchmark, which has models directly control a real robot arm across eight tasks. Argon placed third with 55 out of 100 points, behind GPT-6 Astra at 69 and GPT-6.1 Sol at 65. These measure different things: answering physical-reasoning questions is not the same test as controlling hardware.

Practice worlds, guided actions and feedback

One Google demonstration starts with a photo of a table covered in party props. Argon builds a virtual replica, including a soft plush cap, devises 77 tasks and generates more than 4,500 training episodes for a robot-control model. Google reports improvements when that control model is tested on a real robot, though it does not disclose the number of attempts behind the success rate.

The inputs need not be photographs. Google also shows Argon turning a pen sketch into a simulation where a humanoid calculates a throw to topple six boxes without falling. That example demonstrates simulated motion, not a physical robot completing the throw.

Human guidance remains substantial. In a cup-righting example described as zero-shot—without prior training for that task—Argon receives a camera image with a centimeter grid and ready-made commands such as grasping and lowering until contact. Simulation prompts also contain long lists of forbidden shortcuts. Google did not publish failure rates for individual demonstrations.

Feedback is another part of the workflow. In a robot-arm demonstration, Argon’s prompt instructs it to inspect logs and camera images after failures, identify a cause and retry with a verified fix. Across three iterations, Google shows execution time falling from 58.9 to 17.4 seconds, with the Lego brick landing centrally in a box on the final attempt.

Argon also evaluates training footage. In Google’s origami example, it divides a dual-arm robot’s folding sequence into steps and flags pauses and failed grips. Those segments receive less weight during training. In footage of a person sorting rings, it identifies where progress advances, stalls or reverses.

Paid access follows the wider release

Google opened Argon to developers, businesses and paying consumers on October 13. Google AI Ultra subscribers can access it in AI Studio and Antigravity; developers can also use the paid Gemini API. PYMNTS reports introductory API pricing of $2 per million input tokens and $10 per million output tokens, doubling after November 30. Tokens are the small chunks of text used to meter usage.