Atlas presented as a unified world model for generation and reconstruction
In a CoRL 2026 presentation, World Labs and Stanford researchers introduced Atlas, a from-scratch autoregressive foundation model designed to combine generation, 3D reconstruction, and simulation. According to a Lei Feng Wang report, Atlas accepts text, images, video, depth, graybox, and 3D-structure inputs to produce pixel frames and synthesize new viewpoints. The team demonstrated a distilled version running at 15 fps and described ongoing tests of Atlas in robot simulation. The speakers said Atlas would soon open worldwide, but the report gave no public-release date or availability details.
Atlas is presented as a single model spanning sparse and dense reconstruction, view-consistent generation, and robot simulation, capabilities the speakers described as central to world models for robotics.
导语:时隔两年,World Labs 的工作已经完全不同了。

时隔两年,World Labs 的工作已经完全不同了。
作者丨齐铖湧
编辑丨岑 峰
2026年11月,CoRL(机器人学习顶会)在奥斯汀开幕。