Gemini 4 Argon for Robotics

Google has detailed how its Gemini 4 Argon model can be applied to robotics, in a developer blog post dated October 13, 2026, by principal software engineer Kanishka Rao. The company calls Argon a general-purpose physical reasoning model and says it scored 87% on the Embodied Reasoning (ERQA) benchmark, which it claims beats every other frontier model.

The post presents example prompts across four areas. In robot control, Argon directed a Boston Dynamics Spot to pick up soda cans and step on a foot-pedal bin, rehearsing first in a MuJoCo digital twin. For converting reality to simulation, Google says it can build calibrated environments from robot camera feeds, first-person human video or a paper sketch. On physical understanding, it labeled a 136-second origami-folding video from an ALOHA 2 robot into 6 subtasks and 17 primitives, flagging fumbles and retries.

The most concrete results concern self-improvement. Starting from one tabletop photo, Argon reportedly generated more than 4,500 simulated episodes that raised a vision-language-action model's real-world success rate from 56% to 75% and cut completion time from 35 to 24 seconds. On a small SO-101 arm, a pick-and-place task dropped from 58.9 seconds to 17.4 over three iterations. Google also says Argon reproduced robotics papers as 68 browser-based sandboxes.

Google frames the work as a step toward "physical AGI." The model, gemini-4-argon, is available through the paid Gemini API tier, and to Google AI Ultra subscribers in AI Studio and Google Antigravity. The figures are Google's own demonstrations, and the post offers no independent verification.