MosaicThinker: On-Device Visual Spatial Reasoning for Embodied AI via Iterative Construction of Space Representation
Haoming Wang & Qiyao Xue & Weichen Liu & Wei GaoDepartment of Electrical and Computer EngineeringUniversity of PittsburghPittsburgh, PA 15261, USA {haw200, qix63, weichenliu, weigao}@pitt.edu · 2026
Read MosaicThinker: On-Device Visual Spatial Reasoning for Embodied AI via Iterative Construction of Space Representation by Haoming Wang & Qiyao Xue & Weichen Liu & Wei GaoDepartment of Electrical and Computer EngineeringUniversity of PittsburghPittsburgh, PA 15261, USA {haw200, qix63, weichenliu, weigao}@pitt.edu (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-07082.txt.
Description
Haoming Wang and three co-authors strengthen spatial reasoning for small vision-language models on device. Embodied AI is moving past object detection toward robot manipulation and action planning. Those tasks need visual spatial reasoning from video inputs, to grasp spatial links between objects and steer device acts. Yet current vision-language models reason weakly about space, lacking 3D knowledge, most of all when tasks span complex relations across many video frames.
Their method, MosaicThinker, is an inference-time computing trick for on-device embodied AI. It gathers fragmented spatial clues from many frames into one global semantic map of space. It then guides the small vision-language model's spatial reasoning over that map with a visual prompt. Tests show it lifts accuracy a great deal on cross-frame spatial reasoning, across task types and complexity levels, on resource-constrained embodied devices.
Catalog record
| CREATOR | Haoming Wang & Qiyao Xue & Weichen Liu & Wei GaoDepartment of Electrical and Computer EngineeringUniversity of PittsburghPittsburgh, PA 15261, USA {haw200, qix63, weichenliu, weigao}@pitt.edu |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 67 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-07082