Reproduce “FlowBot3D: Learning 3D Articulation Flow to Manipulate Articulated Objects” 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
Ben Eisner, Harry Zhang, David Held
FlowBot3D introduces 3D Articulation Flow (3DAF), a hierarchy-free representation for robot manipulation of articulated objects. This approach enables robust generalization to unseen object categories and demonstrates successful sim-to-real transfer with a single model trained entirely in simulation at Carnegie Mellon University.
Key insights
- 3D Articulation Flow (3DAF) provides a powerful hierarchy-free and implicit representation of articulation, overcoming limitations of explicit kinematic models.
- Separating perception (3DAF prediction) from analytical motion planning leads to a more robust and generalizable manipulation policy compared to end-to-end learning.
- Self-supervised training of ArtFlowNet entirely in simulation, coupled with viewpoint augmentation, enables effective sim-to-real transfer without fine-tuning.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.