3D Feature Field

3D feature fields represent 3D scenes as continuous fields of features, enabling richer scene understanding and manipulation than traditional methods. Current research focuses on leveraging these fields for tasks like viewpoint optimization, vision-language navigation, and robotic manipulation, often integrating pre-trained 2D models and employing architectures such as neural radiance fields (NeRFs) and transformers. This approach facilitates improved scene decomposition, semantic segmentation, and object manipulation, with applications ranging from autonomous navigation to more intuitive 3D scene editing and robotic control.