Reproduce “Muninn: Your Trajectory Diffusion Model But Faster” 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
Gokul Puthumanaillam, Hao Jiang, Ruben Hernandez, Jose Fuentes, Paulo Padrao, Leonardo Bobadilla, Melkior Ornik
A training-free caching wrapper called Muninn accelerates diffusion-based trajectory planners for real-time robotics by providing quantifiable guarantees on trajectory fidelity, achieving significant latency reductions (e.g., 4.6x on AntMaze) while maintaining task performance and safety.
Key insights
- Changes in inexpensive internal representations of a diffusion denoiser serve as a reliable, low-cost probe to estimate the stability of a trajectory and potential reuse error.
- The analytical form of common diffusion samplers allows for quantifying how errors in predicted noise propagate through subsequent denoising steps, providing pathwise sensitivity coefficients.
- Split-conformal prediction can be effectively applied to calibrate probabilistic upper bounds on reuse errors, enabling a global guarantee on the final trajectory deviation without distribution assumptions.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.