Google Trains Gemini 4 Argon to Control Robots and Build...

Google Trains Gemini 4 Argon to Control Robots and Build Simulations for Them

Google published a breakdown of Gemini 4 Argon on the Developers Blog — a model for physical reasoning that scored 87% on the ERQA benchmark through pure reasoning without external tools, demonstrated in a demo on the Spot robot, and built a MuJoCo simulation with 4,500+ episodes from a single photo of a workspace, fine-tuning on which raised the success rate of a VLA model on a real robot from 56% to 75%.

Compiled by Sergey KostenchukPublished 2026-10-16Updated 2026-10-16

2026-10-16 Research Google

🤖 Google Trains Gemini 4 Argon to Control Robots and Build Simulations for Them

Google published a breakdown of Argon on the Developers Blog — a model for physical reasoning and robotics. On the ERQA benchmark, it scored 87% through pure reasoning — according to Google, without external tools. In the demo, Argon rehearsed the Spot robot dog stepping on a bucket pedal in MuJoCo, and the real robot repeated it after verification by agent verifiers.

🌍 Argon addresses a bottleneck in robotics — collecting real data. From a single photo of a stand with two Franka FR3s, the model built a calibrated MuJoCo simulation and generated 4,500+ episodes; a VLA model fine-tuned on them raised the success rate from 56% to 75%.

👤 Practitioners are shown a pattern: rehearsing a skill in a digital copy of the scene, then transferring it to hardware with agent verification. Video demo and full breakdown with prompts — in the Google blog.

Source 1: https://developers.googleblog.com/gemini-4-argon-for-robotics/

Sources