Google's Gemini 4 Argon turns one table photo into 4,500 robot training episodes

The gain is 19 percentage points in success rate ^[13], and execution time fell from 35 to 24 seconds, about 31 percent ^[5][14]. Both come from one scene: a table with party props, rebuilt from a single photo ^[4]. For the numbers to transfer, your scene would have to look like that table to the simulator. Your control model would also need training episodes of similar quality.

Volume per task is modest. Google's 77 tasks and more than 4,500 episodes ^[4] work out to at least 58 episodes per task if they were spread evenly ^[15].

The Lego demo shows where the human sits. The prompt told Argon to review its own logs and camera images, find the cause of a failure, and retry after applying a verified fix ^[8]. The retry loop is specified in the prompt. The runs then went from just under a minute to 23 seconds to about 17 seconds, and the 23-second run clipped the edge of the box ^[8].

Google pitches one model for control, world generation and data grading ^[1]. The demos pull in more parts. In the Scrabble demo, Gemini Omni, a separate video world model, generates the clips, while Argon acts as the director: it writes timestamped instructions and checks individual frames ^[7]. It took 33 steps to put five tiles in order ^[7]. In the Spot demo, two extra AI agents verify the foot and gripper before the can drops ^[12].

The Decoder reports an ERQA score of 87 percent for Argon, on 400 questions with no helper tools ^[10]. The available text of the report ends before any comparison scores. In our view, a question-answering score tells you about robot control only if the two track each other. They are different measurements.

Google also says it has built 68 browser sandboxes of published robotics experiments, one prompt per paper, including robot soccer and a humanoid on a skateboard ^[9].