paper/arxiv/2609.00188

Robotics Dataset

ZimaBlue is a framework for learning generalizable World Action Models from large-scale egocentric videos, using a three-stage training curriculum and a Slow-Fast dual-system architecture for real-time control, achieving significant improvements in zero-shot robot manipulation.

Catalog metadata

Repository owner

Xionghao Wu, Yijun Yang, Shiyang Zhou et al.

Source

Research paper

Task

ZimaBlue: Evolving Generalizable World Action Models through Scalable Video Pre-training

Cameras

0

Browse the catalog to compare robots, tasks, sensors, and licenses. Sign in to open protected download and model links.