Gemini 4 Argon 用单张照片生成机器人仿真训练数据
Google 称 Gemini 4 Argon 把一张桌面照片转成仿真环境和 4500+ 训练回合，让真实机器人成功率从 56% 提升到 75%。
Three ways robot builders are getting data, from what I read over the past week: ◉ Human video. Nature Will says its Natura-1 hand control model trained on 1M+ hours of first-person video of people using their hands (company claim, via Forbes). ◉ Simulation. Google says Gemini 4 Argon turned one photo of a real tabletop into a sim twin and 4,500+ episodes, lifting real robot success from 56% to 75% (Google's own numbers). ◉ Scanners. Hesai's new Kosmo combines lidar and cameras to map real places. Its CEO called the shortage "a significant gap in high quality spatial data." Different routes, but all three start from the real world. @vangrid_io is the phone capture route, with provenance on Base