Gemini 4 Argon robotics demos: 76-DoF humanoid sim and 87% on ERQA
A video roundup shows Argon turning a hand-drawn sketch into a 76-DoF humanoid simulation, training a control model on 4,500+ simulated episodes to raise real-world success from 56% to 75%, and scoring 87% on the ERQA physical reasoning benchmark.
Gemini 4 Argon combines visual reasoning and code generation across the robotics development stack. Google’s demos cover four areas: • Robot control through camera feedback and motion APIs. • Building physics simulations from recordings and sketches. • Labeling manipulation videos and estimating task progress. • Generating training data and improving actions through repeated trials. One demo turns a hand-drawn sketch into a 76-DoF humanoid simulation, verifying a balanced throw that knocks down six boxes. Another generates 4,500+ simulated episodes to train a robot control model. Google reports real-world success improving from 56% to 75%, with task time falling from 35s to 24s. Argon also scores 87% on ERQA, a benchmark for physical and spatial reasoning. In the attached clip: real robot camera feeds are on the left; a MuJoCo digital twin reconstructed by Argon is on the right. Argon uses recorded motion and sensor data to iteratively tune the scene to match the camera views and physical contact behavior.