Geometry-aware RL for Manipulation of Varying Shapes and Deformable Objects
Tai Hoang, Huy Le, Philipp Becker, Ngo Anh Vien, Gerhard Neumann · 2025
Read Geometry-aware RL for Manipulation of Varying Shapes and Deformable Objects by Tai Hoang, Huy Le, Philipp Becker, Ngo Anh Vien, Gerhard Neumann (2025) 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-2502-07005.txt.
Description
Hoang and co-authors built a reinforcement learning system that manipulates objects of many shapes, including soft ones like rope and cloth. Such tasks are hard because the robot must model how each object moves and bends, and start and goal poses vary widely. The team represents the whole scene as a graph of connected parts. Different link types stand for the grippers, the object, and their contact. One structure fits rigid and deformable jobs alike.
Their policy network passes messages over this graph while respecting three-dimensional symmetry, a design they call HEPi. On a new benchmark covering insertion of varied objects and multi-arm rope and cloth tasks, it beat transformer-based and simpler graph policies on returns, learning speed, and transfer to unseen objects. The authors argue that building geometry and part differences into the policy is what makes the difference.
Catalog record
| CREATOR | Tai Hoang, Huy Le, Philipp Becker, Ngo Anh Vien, Gerhard Neumann |
|---|---|
| YEAR | 2025 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 74 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2502-07005