Paper summary: Utonia: Toward One Encoder for All Point Clouds

Problem

Point-cloud models are usually trained per domain: driving LiDAR, indoor scans, CAD objects. Could one model learn from all of them?

Method

Utonia trains a single self-supervised point transformer encoder across remote sensing, outdoor LiDAR, indoor RGB-D sequences, object-centric CAD models, and point clouds lifted from RGB-only videos. These domains differ in sensing geometry, density and priors.

Key results

From the abstract:

• a consistent representation space that transfers across domains
• better perception, plus emergent behaviours that only appear with joint training
• conditioning vision-language-action policies on Utonia features improves robotic manipulation, and adding them to VLMs improves spatial reasoning

Limitations

There’s no dedicated limitations section in the extracted text. The authors present it as a first step toward foundation models for sparse 3D data.