Google demonstrates Gemini 4 Argon robotics on Apollo humanoid

Google says its newly released Gemini 4 Argon model helped Apptronik’s Apollo humanoid land seven consecutive cornhole shots after analyzing video of a human playing the game. The internal trial anchors the company’s robotics pitch for Argon, which became globally available to developers and enterprises through its API on October 13.

According to Google’s launch post, Argon watched raw video, mapped the environment’s spatial coordinates and calculated the required whole body joint movements. It then autonomously generated simulation code. Google reports that the resulting work was transferred to Apollo for the physical trial.

The Gemini 4 Argon robotics demonstration targets a specific development bottleneck: building a simulation from an observed physical task before attempting it on hardware. Google calls the workflow “Video-to-Sim-to-Real” and says it compresses weeks of manual simulation modeling into a few hours. The reported result is an internal demonstration of one throwing task, rather than a general measure of robot competence.

Physical reasoning and simulation generation

Google also reports an 87% score on the Embodied Reasoning benchmark, ERQA, which tests understanding of physical and spatial problems in 3D environments. The company says Argon outperforms other frontier models using pure reasoning. That benchmark measures reasoning about physical situations; the Apollo trial supplies the separate hardware demonstration.

The distinctive element in Google’s account is the chain of operations, not simply a humanoid throwing a bag. Argon interpreted a human demonstration, worked out spatial relationships and joint movements, and wrote the simulation code needed for the subsequent robot trial. For robotics developers, that makes the claimed reduction in simulation preparation time the more useful result to scrutinize.

Industrial perception is another application highlighted in the launch. A testimonial published by Google says Argon has been integrated into Roboflow for detection and text extraction, with uses including product inspection, inventory tracking and machine guidance. The testimonial reports that Argon ranked first in its object detection evaluations and led a text benchmark using manufacturing, logistics and aerial images.

Developer access is through the paid Gemini API in AI Studio and Google’s Gemini platform. Google lists a unified input and output token limit of one million tokens, alongside introductory pricing of $2 per million input tokens and $10 per million output tokens. After the introductory period expires on November 30, 2026, those rates rise to $4 and $20 respectively.

Source: blog.google