Gemini 4 Argon Hits 87% on Robotics Reasoning Test — SkimNews
Get the Tech newsletter
Daily tech — startups, AI labs, chips, the launches that shape the next decade. Free.
- Gemini 4 Argon scores 87% on the Embodied Reasoning (ERQA) benchmark, outperforming every other frontier model on physical and spatial problem-solving in 3D environments.
- Argon controls a Boston Dynamics Spot quadruped (12-DoF legs + 6-DoF arm) to step on a foot-pedal trash bin and dispose of soda cans, first building an articulated MuJoCo digital twin to rehearse the motion.
- Argon steers the Gemini Omni World Model to rearrange Scrabble tiles spelling G-R-A-N-O into A-R-G-O-N using timestamped action prompts in a closed visual feedback loop.
- Argon reconstructs MuJoCo digital twins from real robot teleoperation episodes, tuning object geometry, friction, and initial poses to match logged tactile force data through iterative visual reasoning.
- Argon converts egocentric human video into 500 Hz MuJoCo manipulation environments, retargeting hand trajectories onto a 76-DoF bimanual humanoid with 22-DoF five-fingered hands.
- Argon can turn a pen-and-paper sketch of a humanoid throwing at a 3-2-1 box stack into a GPU-accelerated mujoco_warp scene with a throw trajectory that topples 6 of 6 targets in closed-loop physics.
Why it matters: Argon's 87% ERQA score and ability to handle quadruped control, MuJoCo simulation reconstruction, and sketch-to-scene generation within one model positions it as a general-purpose physical reasoning layer across the entire robotics lifecycle — from perception and planning through simulation and deployment.
Ask SkimNews
