Generative AI & Models13 October 2026, 16:01 UTC13 min read

Gemini 4 Argon for Robotics

The release is aimed at accelerating progress toward physical AGI and expanding workflows across the robotics stack. DeepMind claims the model can be applied across the entire robotics lifecycle, from planning to execution.

1 source primary source

Key points

Gemini 4 Argon scores 87% on ERQA benchmark.
Model can control Boston Dynamics Spot quadruped to perform trash bin task.
Argon supports MuJoCo digital twin creation for robotic manipulation.

On the Embodied Reasoning (ERQA) benchmark, Argon scores 87 %, outperforming every other frontier model. The article shows the model controlling a Boston Dynamics Spot quadruped to step on a foot‑pedal trash bin and throw soda cans, and a robotic arm to reorient a mug and place a napkin. Argon also builds MuJoCo digital twins and runs closed‑loop visual‑feedback loops.

The release includes detailed prompts, code‑as‑policy primitives, and a visual‑coordinate harness for manipulation tasks. DeepMind also demonstrates Argon’s ability to generate physically accurate MuJoCo environments, synthesize 76‑DoF humanoid trajectories, and annotate real‑world videos with multi‑granularity hindsight. These capabilities suggest a new toolset for researchers and developers working on embodied AI.

Model page: Gemini 4 Argon →

Full story from Google DeepMind · by Kanishka Rao primary sourceOpen source ↗