Gemini 4 Argon 演示：单张照片生成机器人仿真训练

1313 @SmaeiliVahid原推发布于 2026.10.18

在 X 上打开原推

Google’s Gemini 4 Argon post has a demo I keep thinking about. Starting from one photo of a real tabletop, it builds a simulation twin and runs 4,500+ randomized episodes. Google says real robot success went from 56% to 75%, and completion time from 35s to 24s. Those are Google’s own numbers from its blog. Synthetic data is clearly getting better. But look at where that pipeline starts: a capture of a real place. The post also lists real inputs like robot camera feeds and egocentric human video. That’s the layer @vangrid_io works on: real captures of real places, with provenance you can check.