Gemini 4 Argon taught to control robots: pick-and-place success rose from 56% to 75% in Google's test
Prepared by Sintelligent with the help of SI. Published in SI: . Updated: .
Robots can now be trained and tested with one general Google model instead of writing a separate program for every movement.
Google has released a robotics version of the model: it controls mobile manipulators, builds simulation copies from video of robots and people and from sketches, and reproduces scientific papers in the MuJoCo simulator. According to Google, the success rate of operations on real robots rose from 56% to 75% and completion time fell from 35 to 24 seconds; the model scored 87% on the ERQA physical reasoning test and reproduced 68 robotics papers. Examples include Boston Dynamics Spot, Franka FR3, ALOHA 2 and humanoids with dexterous hands. Access is via the Gemini API (gemini-4-argon), Google AI Studio and Antigravity for Ultra subscribers.
Sources 1