Google DeepMind releases Gemini 4 Argon for robotics

Google DeepMind has introduced Gemini 4 Argon, a Gemini 4 variant built for robotics. The company describes it as a general-purpose physical reasoning model with gains in physical understanding, code generation and visual-semantic reasoning, and says it can be used across the whole robotics lifecycle.

The main result is on the Embodied Reasoning (ERQA) benchmark, which tests whether a model can understand and solve physical and spatial problems in 3D environments. Google says Argon scores 87% "using pure reasoning," outperforming every other frontier model. Google also reports a gain on real robots: a vision-language-action (VLA) policy's success rate rose from 56% to 75% after training on data that Argon generated in simulation. The excerpt available here does not give the full setup for either result.

Most of the post consists of example prompts across four areas:

• Robot control with sim rehearsal: Argon controls a Boston Dynamics Spot quadruped (12-DoF legs plus a 6-DoF arm) to throw soda cans into a foot-pedal trash bin. It first builds an articulated MuJoCo digital twin of the bin to practise the move, and uses a two-agent "verification swarm" to check foot and hand placement.
• Code-as-policies manipulation: Argon sets a fallen mug upright and puts a napkin inside it over several turns, using a wrist camera with a coordinate grid and HTTP motion primitives.
• World-model steering: Argon writes timestamped prompts for Google's Gemini Omni video world model, then checks the rollout frames and refines its prompts until simulated robots arrange Scrabble tiles to spell "ARGON."
• Real-to-sim: Argon builds MuJoCo digital twins from real teleoperation episodes and tunes the assets until simulated contact events match the recorded data.

Google says Argon is available to developers at launch. Pricing and access details were not in the material reviewed.