Paper 2603.03283
Utonia: Toward One Encoder for All Point Clouds
- Published
- Mar 2026
- Research lab
- Independent
- Citations
- 9
- GitHub
- 717 stars
01 In brief
Summary
Utonia introduces a single self-supervised point transformer encoder trained jointly on diverse point cloud domains, including indoor scans, outdoor LiDAR, remote sensing, object CAD, and video-lifted point clouds.
The authors identify three core cross-domain mismatches: inconsistent modality availability, sensitivity to granularity shifts, and bias toward gravity conventions.
They propose three domain-agnostic fixes: Causal Modality Blinding (randomly dropping color/normal channels during pretraining), Perceptual Granularity Rescale (aligning spatial units across domains), and RoPE-enhanced positional hints on granularity-aligned coordinates.
Trained on 250k cross-domain point clouds plus 1M CAD assets, Utonia achieves competitive or superior performance on indoor, outdoor, and object-centric benchmarks compared to prior domain-specific methods like Sonata and Concerto.
Notably, Utonia shows robustness when colors or normals are missing, and exhibits emergent behaviors: improved cross-domain transfer, gravity-aligned scene features with object rotation invariance, and benefits in downstream applications including robotic manipulation (82.1% success rate vs.
80.0% for Concerto), open-world part segmentation, and spatial reasoning in vision-language models.
The work demonstrates that a unified point encoder is feasible and beneficial, paving the way toward foundation models for sparse 3D data.
Future directions include query-based task interfaces, 4D spatial cognition, and scalable next-generation backbones.
The project is supported by Hong Kong RGC and NSFC grants, with code available at https://pointcept.github.io/Utonia.
The paper is authored by researchers from the University of Hong Kong, Chinese University of Hong Kong, and Xiaomi, and is available on arXiv (2603.03283v2).
02 From the paper
Abstract
We dream of a future where point clouds from all domains can come together to shape a single model that benefits them all. Toward this goal, we present Utonia, a first step toward training a single self-supervised point transformer encoder across diverse domains, spanning remote sensing, outdoor LiDAR, indoor RGB-D sequences, object-centric CAD models, and point clouds lifted from RGB-only videos. Despite their distinct sensing geometries, densities, and priors, Utonia learns a consistent representation space that transfers across domains. This unification improves perception capability while revealing intriguing emergent behaviors that arise only when domains are trained jointly. Beyond perception, we observe that Utonia representations can also benefit embodied and multimodal reasoning: conditioning vision-language-action policies on Utonia features improves robotic manipulation, and integrating them into vision-language models yields gains on spatial reasoning. We hope Utonia can serve as a step toward foundation models for sparse 3D data, and support downstream applications in AR/VR, robotics, and autonomous driving.