Google shows Gemini 4 Argon controlling robots and building simulations
Google demonstrated Gemini 4 Argon performing robot tasks and generating simulations, alongside an 87% score on a physical reasoning benchmark.
Google presented robotics demonstrations for Gemini 4 Argon on October 13. The company said the model scored 87% on ERQA, a benchmark assessing reasoning about physical environments and spatial problems.
One demonstration used a Boston Dynamics Spot robot to dispose of cans in a pedal-operated bin after practicing in simulation. In another experiment, Google said simulation-generated training data increased a robot’s task success rate from 56% to 75%, while completion time fell from 35 seconds to 24 seconds. These are company-reported experimental results.
The work connects visual interpretation, simulation generation and physical control within a single model. Google is making Argon available through its paid Gemini API, while Ultra subscribers can access it in Google Antigravity. Developers can now use those services to build on the demonstrated workflows.
The day's biggest AI news in one short email, every morning. Plus an instant heads-up when a reset or a big release lands.
More on Google DeepMind
- Oct 13Google expands Gemini 4 Argon access to developers, businesses and subscribersOfficial
- Oct 13Cursor adds Gemini 4 Argon with a 52.8% CursorBench scoreOfficial
- Oct 13Google details Gemini 4 Argon safeguards as broader rollout beginsOfficial
- Oct 13Gemini CLI Release v0.64.0 is outUpdate
- Oct 6Gemini CLI Release v0.63.0 is outUpdate
- Oct 1From deepfakes to DNA: the science of watermarking AIOfficial video