Google pitches Gemini 4 Argon as a robotics tool for simulation and data, not as the controller

Google published “Gemini 4 Argon for Robotics” on October 13, 2026, as part of the launch of its Gemini 4 Argon frontier model. The post shows the general model building physics simulations, labelling robot data and, in a few demos, directing robots. Google says Argon scores 87% on the ERQA embodied-reasoning benchmark, “outperforming every other frontier model”, and that more than 4,500 simulated training episodes it generated raised a physical robot’s success rate on one task from 56% to 75%.

Why it matters: Google is not offering Argon as a robot’s brain. It is offering it as a tool for the parts of robot development that eat engineering time, while its robot-specific models keep the control loop.

What does Google say Argon can do for robots?

The post, by Google principal software engineer Kanishka Rao, groups its demos into four areas. In control, Argon directs a Boston Dynamics Spot with a 6-DoF arm to step on a foot-pedal bin after practising on a MuJoCo digital twin of the bin. In a zero-shot demo, it has a simple robot arm stand a fallen mug upright and drop a napkin into it, working over several turns through a coordinate harness and preset action primitives.

In Real2Sim, Argon rebuilds a MuJoCo scene from a real teleoperation episode and tunes assets and contacts against the recording. It also turns human hand video into motion for a simulated bimanual humanoid, and a hand-drawn sketch into a MuJoCo scene with an Apollo 2 humanoid. For data work, it writes a four-level hindsight annotation for a 136-second ALOHA 2 origami video and down-weights fumbles to 0.15–0.25x for behaviour cloning. In self-improvement, it cut an SO-101 arm’s pick-and-place time from 58.9 seconds to 17.4 seconds over three iterations.

Google calls the prompts “representative”. It does not say how many trials sit behind the 56% and 75% figures, and the ERQA comparison names no rival models.

Why isn’t Argon the robot controller?

Google’s own documentation draws the line. The Gemini API model page lists the Live API as unsupported for gemini-4-argon. The robotics developer docs use only Gemini Robotics-ER 2 models, including a streaming endpoint built on the Live API for real-time use. Argon does not appear in them. Below ER 2 sits the Gemini Robotics 2 vision-language-action model, which turns camera input into motor commands.

So the stack Google is describing has three layers. Argon works offline, before a robot moves. ER 2 plans while it moves. The VLA executes. Argon’s best-supported robotics result is about data rather than control: simulated episodes that improved a real policy. That targets the same bottleneck that has pushed companies toward sim-to-real training and pushed China to build teleoperation data centres. A general model that can build and calibrate the simulation is a cheaper way into that work than a team of simulation engineers, if the gains hold beyond one tabletop.