OMEGALAB Omnimodal Multi-Embodiment Generalist Agents HKUST(GZ)

MULTIMODAL PERCEPTION · 2026

UniT

Unified Geometry Learning with Group Autoregressive Transformer

  1. 1Intelligent Transportation Thrust, Systems Hub, HKUST(GZ)
  2. 2National Key Laboratory of Human-Machine Hybrid Augmented Intelligence, Xi'an Jiaotong University
  3. 3Applied Science, Amazon.com, Inc.
  4. †Corresponding authors
Paper Code Live demo BibTeX

OVERVIEW

One model for geometry perception

UniT is a unified feed-forward model that reformulates a wide range of geometry perception capabilities into a single framework, covering diverse view configurations, modality combinations, metric-scale perception, and long-horizon scalability. It supports both online and offline inference over an arbitrary number of views, flexibly incorporates auxiliary modalities such as camera parameters and depth maps, recovers geometry at metric scale in physical units, and maintains bounded complexity over long horizons in in-the-wild environments.

UniT overview: view configurations, auxiliary modalities and the group autoregressive transformer that produces metric-scale geometry.
Overview of UniT.

EXAMPLES

Reconstructions in the wild

Point clouds reconstructed by UniT, from campus scenes at HKUST(GZ) to public driving and synthetic sequences. Hover or tap a clip to play it.

  • HKUST(GZ) · INTR
  • HKUST(GZ) · Toy
  • HKUST(GZ) · Red Bird
  • Drift
  • GTA SfM
  • KITTI

Explore the point clouds interactively and measure real-world distances on the original project page

LIVE DEMO

Run UniT on your own images

The Hugging Face Space runs the released model in your browser, no setup required.

Open the demo

RESULTS

Evaluated across tasks and settings

Compared with VGGT, π3, MapAnything, DepthAnything3, CUT3R and StreamVGGT under both scale-invariant and metric-scale settings.

Benchmarks
10
Tasks
7
Baselines
6

Scene-level 7-Scenes and NRGBD, and object-centric DTU.

Table 3: multi-view reconstruction results on 7-Scenes, NRGBD and DTU.
Figure 8: qualitative multi-view reconstructions compared with ground truth, shown without alignment or filtering.
Qualitative multi-view reconstruction. Point clouds are shown raw, without alignment or filtering; each row uses a consistent scene scale.

CITATION

@misc{wang2026unit,
  title={UniT: Unified Geometry Learning with Group Autoregressive Transformer},
  author={Haotian Wang and Yusong Huang and Zhaonian Kuang and Hongliang Lu and Xinhu Zheng and Meng Yang and Gang Hua},
  year={2026},
  eprint={2605.21131},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2605.21131},
}