Reproduce “Learning Part-Aware Dense 3D Feature Field for Generalizable Articulated Object Manipulation” 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

Yue Chen, Muqing Jiang, Kaifeng Zheng, Jiaqi Liang, Chenrui Tie, Haoran Lu, Ruihai Wu, Hao Dong

Researchers from Peking University, Beijing Institute of Technology, and the National University of Singapore developed a Part-Aware 3D Feature Field (PA3FF) and an associated Part-Aware Diffusion Policy (PADP) for generalizable articulated object manipulation. The method leverages dense, semantic, part-aware 3D features to achieve a 28.79% average success rate on the PartInstruct benchmark in simulation and 58.75% mean success in diverse real-world tasks, outperforming prior approaches.

Key insights

  • Providing robots with a 3D-native, dense, semantically grounded, and explicitly part-aware feature representation is critical for generalizable articulated object manipulation.
  • Leveraging and adapting strong 3D geometric priors (e.g., from Point Transformer V3) in combination with a part-aware contrastive learning framework effectively addresses limitations of 2D-lifted features.
  • Integrating these refined, part-aware 3D features within an imitation learning framework like diffusion policies leads to enhanced generalization and robustness across unseen objects, states, and environments.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.