Reproduce “Object-centric Task Representation and Transfer using Diffused Orientation Fields” 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
Cem Bilaloglu, Tobias Löw, Sylvain Calinon
Researchers from the Idiap Research Institute and EPFL present Diffused Orientation Fields (DOF), a method for creating continuous, geometry-aware local reference frames crucial for robotic manipulation on curved objects. This approach enables the definition of shape-invariant actions, facilitating successful transfer of contact-rich skills across diverse object geometries and demonstrating improved learning performance and zero-shot generalization in reinforcement learning.
Key insights
- Diffusing a scalar field, initialized by sparse semantic keypoints, across an object's surface effectively generates a smooth, geometry-aware direction field within the object's tangent planes.
- The Walk on Spheres algorithm provides an efficient, grid-free method to extend these surface-defined orientation fields into the entire 3D workspace, making them accessible for any robot pose without full workspace discretization.
- Representing manipulation tasks within these smoothly varying local reference frames inherently decouples task specification from specific object geometries, allowing robust and generalized skill transfer with only keypoint correspondences.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.