Reproduce “Object-Oriented Dynamics Learning through Multi-Level Abstraction” 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
Guangxiang Zhu, Jianhao Wang, Zhizhou Ren, Zichuan Lin, Chongjie Zhang
A Multi-level Abstraction Object-oriented Predictor (MAOP) framework learns action-conditioned dynamics from raw pixels, achieving zero-shot generalization and high sample efficiency across various multi-object environments. This enables robust model-based planning and provides interpretable disentangled object representations.
Key insights
- A multi-level abstraction architecture, inspired by cognitive principles, effectively decomposes complex dynamics learning into progressively finer-grained, easier-to-solve sub-problems.
- Guiding the object-oriented dynamics learning with proposals from an initial motion detection and dynamic instance segmentation level improves robustness and sample efficiency.
- Learning disentangled object representations, including their relations and temporal effects, is crucial for achieving zero-shot generalization to novel environmental configurations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.