Reproduce “Generalizable Motion Planning via Operator Learning” 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
Sharath Matada, Luke Bhan, Yuanyuan Shi, Nikolay Atanasov
Researchers at the University of California, San Diego, developed a Planning Neural Operator (PNO) that learns to solve the Eikonal Partial Differential Equation for motion planning, generalizing across various obstacle geometries, resolutions, and goal positions without retraining. This PNO achieves up to a 10x speedup over numerical solvers in high-resolution 2D maps and a 2-3x speedup in 3D environments while maintaining high accuracy and demonstrating zero-shot super-resolution capabilities.
Key insights
- Learning an infinite-dimensional solution operator for the Eikonal PDE allows generalization across entire families of environments and resolutions, unlike traditional neural networks.
- The inherent resolution invariance of Fourier Neural Operators enables PNO to be trained on coarse data and predict accurately on significantly finer resolutions without retraining.
- Explicitly encoding obstacle information and enforcing triangle inequality through architectural biases (Deepnorm layer) are crucial for robust generalization across environment geometries and goal positions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.