ECD / README.md
nielsr's picture
nielsr HF Staff
Add model card, pipeline tag, and sample usage
6f58a37 verified
|
Raw
History Blame
2.38 kB
metadata
license: apache-2.0
pipeline_tag: reinforcement-learning

Energy-based Compositional Diffusion Planning (ECD)

This repository contains checkpoints for ECD (Energy-based Compositional Diffuser), presented in the paper Energy-based Compositional Diffusion Planning (ICML 2026).

Introduction

Compositional diffusion planners aim to solve long-horizon robotic tasks using short training trajectories. ECD (Energy-based Compositional Diffuser) is an inference-only framework that formulates the global trajectory as the minimizer of the sum of local bridge potentials. Instead of stitching local chunk predictions heuristically, ECD defines a single global energy function over all chunks, using its negative gradient to guide the denoising process, ensuring conservative score fields and consistent global modes.

Usage: Using ECD as a Plug-in

Because ECD operates entirely at inference time, it can wrap any pretrained short-horizon diffusion denoiser. Below is an example of how to instantiate the CompositionalPolicy and set the inference type to ecd_chunk:

from ecd.policy import CompositionalPolicy

policy = CompositionalPolicy(
    diffusion_model=denoiser,                 # Your short-horizon chunk denoiser (see ecd/planner.py)
    normalizer=normalizer, 
    ev_n_comp=N,
    ev_cp_infer_t_type="ecd_chunk",           # Change to "interleave" for the standard CD baseline
    ecd_config=dict(
        rank_type="overlap",                  # Map-free candidate ranker
        base_scale=0.15, 
        react_scale=0.10,                     # Interior update / boundary-reaction strength
        markov_type="laplacian", 
        chunk_react_type="markov",
    ),
)

# Generate a long-horizon plan
plan = policy.plan(start_xy, goal_xy, b_s=40) 

Citation

@inproceedings{sun2026ecd,
  title     = {Energy-based Compositional Diffusion Planning},
  author    = {Sun, Tao and Mishra, Utkarsh A. and Lu, Jiaxin and Xu, Danfei and Armeni, Iro},
  booktitle = {Proceedings of the 43rd International Conference on Machine Learning (ICML)},
  series    = {PMLR},
  volume    = {306},
  year      = {2026}
}