The world model from World Labs allows for the reconstruction of physical scenes into simulated environments that support autonomous navigation. The Atlas model was demonstrated by mapping Nvidia's Voyager headquarters using only 32 input images to allow users to explore the park in real time. By integrating text, images, video, and 3D data into a shared spatial context, the system creates new views and reconstructs scenes with pixel perfect camera control.

World Labs pretrained the Atlas model from scratch using Nvidia Blackwell GPUs and built it on an open platform. The company also revealed zero shot fully autonomous drone flight powered by scaled embodied foundation models that unify an internal world model with embodied reasoning and direct control.