Gemini 4 Argon 机器人开发栈演示：ERQA 87%
视频汇总 Argon 在机器人领域的四项能力：手绘草图生成 76 自由度人形仿真、4500+ 仿真回合训练控制模型使真机成功率 56%→75%，ERQA 物理空间推理得 87%。
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.