A new result that I am very excited about: A self-supervised visuospatial representation with better 3D awareness.
If you have RGB-D inputs (robots usually have), our model makes use of the depth to give you better semantic and geometric features than DINO.
📄 arxiv.org/abs/2608.27226
