Google DeepMind launches Gemini 4 Argon for robotics
A physical-reasoning model that writes its own simulations before touching hardware
The facts
- Argon scores 87% on the Embodied Reasoning (ERQA) benchmark using pure reasoning, which Google DeepMind says outperforms every other frontier model
- The model covers four robotics workflows: robot control, Real2Sim simulation reconstruction, physical understanding, and research/self-improvement
- Demonstrated controlling a Boston Dynamics Spot quadruped (12-DoF legs plus 6-DoF arm) to practice a foot-pedal trash-bin task in MuJoCo before executing it physically
- Can reconstruct physically calibrated simulation twins from robot camera feeds, egocentric human video, or hand-drawn paper sketches
- Shown steering the Gemini Omni video world model through iterative timestamped action prompts to rearrange Scrabble tiles spelling A-R-G-O-N
- Can retarget egocentric human video onto a 76-DoF bimanual humanoid with 22-DoF five-fingered hands inside MuJoCo contact physics at 500 Hz
Google DeepMind has released Gemini 4 Argon, a version of its Gemini line built for physical reasoning and robot control. The company says Argon marks a step change in understanding physical environments, generating code, and reasoning about visual and semantic information together. It scores 87% on the Embodied Reasoning benchmark, which tests spatial problem-solving in 3D environments, ahead of every other frontier model it compares against.
Rather than a narrow controller, Argon is pitched as a general-purpose model spanning the whole robotics lifecycle. Google DeepMind demonstrates it directing a Boston Dynamics Spot quadruped through a multi-step trash-disposal task. It first builds a MuJoCo digital twin of a foot-pedal bin, to rehearse the motion before the robot steps on the real pedal. In a separate example, Argon operates a simple robotic arm through a calibrated wrist-camera coordinate system. It issues curl commands turn by turn to reorient a fallen mug and tuck a napkin inside it.
The model's Real2Sim capability turns raw observation into physics into something simulatable. It reconstructs a calibrated MuJoCo scene from a robot's own camera feed, from egocentric human video, or from a hand-drawn sketch. It then checks that simulated contact forces and camera renders line up with the real recordings. In one case it retargets human hand and wrist motion from first-person video onto a 76-DoF bimanual humanoid with 22-DoF five-fingered hands. It steps the physics at 500 Hz to verify force closure and stable placement. Argon also steers Google's Omni video world model through an iterative visual-feedback loop, writing and revising timestamped action prompts until Scrabble tiles land spelling out its own name.
Google DeepMind frames all of this as groundwork for what it calls physical AGI, artificial general intelligence operating in the physical world. But it gives no release timeline, pricing, or availability details for Argon beyond the blog post and sample prompts.
Why it matters
The real story is that Argon writes its own simulation code and verification loops before acting. That collapses a step robotics teams usually build by hand. For teams running sim-to-real pipelines, the bottleneck shifts from engineering MuJoCo scenes to supplying good video and sketch data for Argon to work from.
Primary source
Google DeepMind — deepmind.googleWritten from the primary source above by The Neural Net’s automated desk (claude-sonnet-5, theneuralnet/pipeline@1) without human review. How this works. Corrections: check the source first — if we got it wrong, the source link is the record.