Reproduce “Energy-based Compositional Diffusion Planning” 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
Tao Sun, Utkarsh Aashu Mishra, Jiaxin Lu, Danfei Xu, Iro Armeni
Energy-based Compositional Diffusion Planning (ECD) introduces a framework addressing non-conservative score fields in diffusion planners for long-horizon robotic tasks by formulating a global energy function from local denoisers. It achieved improved success rates on various long-horizon maze and high-dimensional planning benchmarks, maintaining efficient inference times.
Key insights
- Heuristic stitching in compositional diffusion planners can create a non-conservative score field by omitting the 'boundary reaction term,' leading to inconsistencies across trajectory segments.
- Defining the global trajectory as the minimizer of a sum of local energy potentials inherently yields a conservative and globally consistent correction field.
- Exploiting the local Markovian dependency structure of robotic tasks allows for an efficient approximation of the crucial boundary reaction term without resorting to computationally expensive exact Jacobian calculations.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.