Utonia: Toward One Encoder for All Point Clouds — Unified Self-Supervised 3D Point Transformer

Summary

Utonia is a unified self-supervised point transformer encoder developed through a collaboration among the University of Hong Kong, the Chinese University of Hong Kong, and Xiaomi. It addresses the 'dialect problem' of 3D point clouds, where indoor RGB-D data is dense and fine-grained, outdoor LiDAR data is sparse and wide-ranging, and CAD models use entirely different coordinate systems—previously requiring separately trained encoders for each data type. Utonia is claimed to be the first cross-domain self-supervised point transformer encoder, jointly trained on heterogeneous point cloud domains including remote sensing data, outdoor LiDAR, indoor RGB-D sequences, object-centric CAD models, and point clouds extracted from RGB video. The authors report emergent behaviors that appear only with multi-domain joint training, such as enhanced robotic manipulation capabilities and improved vision-language-action policy reasoning. Demonstrated applications include object separation and occlusion robustness in robotic manipulation, open-world 3D segmentation, and improved spatial reasoning for vision-language models. Paper: https://arxiv.org/abs/2603.03283; Project page: https://pointcept.github.io/Utonia.