Gemini 4 Argon for Robotics

Gemini 4 Argon is a multimodal model that advances physical understanding, code generation, and visual-semantic reasoning to support a wide range of robotics workflows. It achieves top performance on the Embodied Reasoning (ERQA) benchmark (87%) and is presented as a general-purpose physical reasoning engine that can be applied across robot control, real-to-sim reconstruction, detailed physical understanding, and research-driven self-improvement. Key design themes are closed-loop visual reasoning, iterative verification, and the ability to generate actionable control plans and simulation assets.

Representative demonstrations show how Argon constructs and uses simulation twins and action primitives to solve real tasks: it builds a MuJoCo twin of a foot-pedal trash bin so a Boston Dynamics Spot practices stepping and dropping cans with a 2-agent foot/hand verification swarm; it performs zero-shot arm control to reorient a mug and place a napkin using a visual coordinate harness and curl-based motion primitives (robot.look_at, grasp, place_until_contact, etc.); it steers a Gemini Omni video world model in timestamped action beats to rearrange Scrabble tiles; and it converts robot teleop episodes, egocentric human videos, or pen-and-paper sketches into calibrated 500 Hz MuJoCo scenes (retargeting to a 76-DoF humanoid with 22-DoF hands) while enforcing visual and physical match objectives and forbidding post-hoc sim “cheats.”