Google Wants Gemini 4 Argon to Be the Only AI Model Needed for Robots
谷歌想让Gemini 4 Argon成为机器人唯一需要的AI模型
Google DeepMind has introduced the general-purpose physical reasoning model Gemini 4 Argon, capable of handling robot control, simulation environment setup, training data quality inspection, and experiment reproducibility. Through simulation training and video generation, Argon demonstrated improved execution efficiency and success rates across multiple tasks. In the ERQA benchmark test, Argon scored 87%, outperforming other models, but showed slightly lower performance in the AA-Robotics benchmark.
Google's Gemini 4 Argon model demonstrates significant progress in integrating simulation, control, and evaluation for robotics tasks, achieving improved performance in real-world scenarios through synthetic data and virtual environments. However, its performance on independent benchmarks like AA-Robotics suggests room for improvement compared to other models.
Source: RoboSpeak — WeChat · Read original article ↗
English translation unavailable. Choose the original article to read it.
What the source reports
Publisher-reported claims, with original evidence. These results have not been independently verified by RoboSignal.
Reported numbers
execution time
24 seconds
Source excerpts and review record
Automatically extracted; no manual editorial approval recorded.
在真正踩下踏板之前， Gemini 4 Argon 先构建出垃圾桶的虚拟副本， 机器人在仿真环境中练习如何用三条腿保持平衡；之后才在真实硬件上尝试这项任务。
Open source S9
在真实机器人上测试时， 成功率从 56% 提升到 75%，执行时间从 35 秒降至 24 秒。
Open source S14
仅凭一张照片，Argon 生成了 4500 多个多样化的仿真片段，使真实机器人的成功率从 56% 提升至 75%。
Open source S15
Source：RoboSpeak — WeChat · mp.weixin.qq.com