Latent Geometry Beyond Search: Amortizing Planning in World Models
Hoang Nguyen, Xiaohao XuXiaonan HuangDepartment of Robotics, University of Michigan, Ann Arbor · 2026
Read Latent Geometry Beyond Search: Amortizing Planning in World Models by Hoang Nguyen, Xiaohao XuXiaonan HuangDepartment of Robotics, University of Michigan, Ann Arbor (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-2605-08732.txt.
Description
Hoang Nguyen, Xiaohao Xu, and Xiaonan Huang replace slow search based planning in learned world models with one fast learned mapping. Vision based world models compress scenes into smooth latent spaces, yet goal planning there usually needs costly iterative search. In a pretrained model whose latent geometry is kept smooth and uniform, they argue planning structure already sits inside the representation itself.
Their goal conditioned inverse dynamics model reads current state, goal state, and time left, then outputs the next action directly. Across navigation, manipulation, and control tests it matches or beats search in nearly all settings at roughly one hundredth of the cost. Structured latent spaces can shift planning from online search to learned inference.
Catalog record
| CREATOR | Hoang Nguyen, Xiaohao XuXiaonan HuangDepartment of Robotics, University of Michigan, Ann Arbor |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 78 KB |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2605.08732