INSAIT

“Rather than inflating task horizon through unstructured scene changes, our benchmark structures long-horizon manipulation across eight spatial and logical task categories”