Gemini 4 Argon for Robotics

Gemini 4 Argon represents a significant advancement in physical reasoning, code generation, and visual-semantic understanding, enabling new capabilities across the robotics pipeline. The model demonstrates its strength on the Embodied Reasoning (ERQA) benchmark, achieving an 87% score through pure reasoning, surpassing all other frontier models in solving 3D physical and spatial problems. This general-purpose model is designed to support the entire robotics lifecycle, from simulation to real-world deployment, by integrating precise physical understanding with actionable insights.

Argon’s capabilities extend to controlling real robots, such as a Boston Dynamics Spot quadruped with a 6-DoF arm, to perform tasks like stepping on a foot-pedal trash bin to dispose of soda cans. It also creates detailed MuJoCo digital twins of objects, enabling robots to practice movements in simulation before execution. The model can manipulate simple robotic arms to upright fallen objects and handle dexterous tasks, such as rearranging Scrabble tiles into a specific order using a closed visual loop and iterative refinement of action prompts.

Beyond physical control, Argon enhances simulation workflows by generating accurate digital twins from real-world data, including egocentric human videos of manipulation tasks. It identifies task-relevant objects, refines meshes for concave containers, and retargets human hand trajectories onto complex humanoid robots with 76-DoF and 22-DoF five-fingered hands. Argon can also construct GPU-accelerated simulation scenes from sketches, synthesize balanced trajectories, and verify outcomes in closed-loop physics, demonstrating its versatility in bridging coding, perception, and robotics.

The model’s applications include improving data generation for vision-language-action models, where it autonomously builds calibrated simulation twins from a single RGB photo and synthesizes thousands of domain-randomized episodes. This process boosts real-world robot success rates from 56% to 75% and reduces task completion time from 35 seconds to 24 seconds. Argon also enables self-improvement in robotic arms by iteratively refining execution speed and placement accuracy through trial and self-reflection, highlighting its potential to accelerate progress toward physical AGI.