Reproduce “Manipulating Elasto-Plastic Objects With 3D Occupancy and Learning-Based Predictive Control” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Zhen Zhang, Xiangyu Chu, Yunxi Tang, Lulu Zhao, Jing Huang, Zhongliang Jiang, K. W. Samuel Au
Researchers from The Chinese University of Hong Kong and collaborating institutions developed a framework for robotic manipulation of elasto-plastic objects, leveraging a dense 3D occupancy representation and a hybrid 3D CNN-GNN dynamics model. This approach, enhanced by a shape-based action initialization for predictive control, achieved robust shaping of complex 3D forms, with deformation predictions exhibiting significantly lower error rates (e.g., EMD 22.8e-3) compared to prior methods and results comparable to human manipulation in real-world tasks.
Key insights
- Dense 3D occupancy provides a superior and inherently volumetric state representation for deformable objects, effectively capturing both internal and external geometry despite occlusions.
- A hybrid 3D CNN-GNN architecture effectively models the complex and irreversible dynamics of elasto-plastic materials by integrating multi-scale spatial features with graph-based relational learning.
- Geometrically informed action initialization significantly improves the efficiency and consistency of model predictive control, particularly for tasks involving irreversible plastic deformations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.