Google DeepMind unveils Gemini 4 Argon for robotics with 87% score on Embodied Reasoning benchmark

Google DeepMind introduced Gemini 4 Argon, a general-purpose physical reasoning model designed to operate across the robotics lifecycle. Argon scored 87% on the Embodied Reasoning (ERQA) benchmark, which evaluates a model’s ability to understand and solve physical, spatial problems in 3D environments, outperforming other frontier models using pure reasoning.

Argon’s capabilities span four areas: robot control, Real2Sim reconstruction, physical understanding, and research/self-improvement. In robot control, Argon demonstrates mobile manipulation with a Boston Dynamics Spot quadruped, performing a foot-pedal trash bin task using a MuJoCo digital twin for practice before real-world execution.

The model also enables zero-shot control of a robotic arm to upright a fallen mug and place a crumpled napkin inside, using a visual coordinate harness and action primitives over multiple turns. Argon steers robots in a closed visual loop within the Gemini Omni World Model, rearranging Scrabble tiles into the sequence A-R-G-O-N through iterative, timestamped action prompts.

In Real2Sim workflows, Argon reconstructs physically calibrated simulation twins from robot camera feeds, egocentric human video, or hand-drawn sketches, tuning assets to match real-world contact dynamics. For physical understanding, Argon performs agentic multimodal hindsight labeling, including dense value functions and progress estimation, while supporting action improvement via simulation and iterative self-improvement on hardware.