This project aims to develop a system that can perform open-world understanding based on pretrained large vision-language foundation models. Open-world understanding aims to perform semantic understanding of the real world on a very large set of open object classes. We will investigate strategies to distill open-world knowledge from existing large vision-language foundation models and implement open-world semantic perception models. The system is expected to perform a semantic perception of more than 1000 classes from various real data distributions.