Gemini 4 Argon for robotics
On 13 October 2026 Google used its developers' blog to show how to use Gemini 4 Argon in robotics, from controlling a robot and building simulation twins to estimating progress and self-improvement. On the Embodied Reasoning benchmark (ERQA) Google says the model scores 87%, outperforming every other frontier model; training in simulation raised the real robot's success rate from 56% to 75%.
Why it matters
Google presents a general-purpose model as one that works across a robot's whole path: it sees, builds a physics simulation of a scene from a photo or video, and writes the control code that is then used to collect training data. This is not a new model but a list of uses of an existing one; the figures are Google about Google, and the post consists mostly of example prompts.
What is in the post. The author is Kanishka Rao. The post says Argon marks a step change in physical understanding, code generation and visual-semantic reasoning. On ERQA, which tests understanding and solving of physical, spatial problems in 3D, Argon scores 87% "using pure reasoning, outperforming every other frontier model". The post shows representative prompts in four areas. Robot control: a Boston Dynamics Spot quadruped learns to step on a foot-pedal bin after practising in a MuJoCo digital twin; zero-shot control of a simple arm; closed-loop steering of the Gemini Omni video world model. Real2Sim: building calibrated simulation twins from robot camera feeds, from egocentric human video (retargeting the hand trajectory onto a 76-DoF bimanual humanoid with 22-DoF five-fingered hands) and from a hand-drawn sketch. Physical understanding: a dense value function and progress estimate over a 60-second episode (28 segments). Research and self-improvement: from a single photo of a tabletop Argon builds a MuJoCo twin of two Franka FR3 arms, defines a curriculum of 77 tasks and collects more than 4,500 episodes in simulation; for generalist vision-language-action models the post says this lifted the real-world success rate on an evaluation benchmark from 56% to 75% and cut completion time from 35 s to 24 s. Separately, an SO-101 arm improves its own pick-and-place (58.9, 22.9 and 17.4 s by the video captions), and 68 robotics papers are reproduced as browser labs. Availability. Google AI Studio and Google Antigravity for Google AI Ultra subscribers; the Gemini API on its paid tier, model gemini-4-argon. What the record does not claim. The videos were not played; the benchmark behind the 56% and 75% and the vision-language-action model are not named in the post; the 87% on ERQA and the comparison with other models are Google's words. No independent check was read.