Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Planning Using Schrödinger Bridge Diffusion Models

Adarsh Srivastava

TL;DR

This paper investigates speeding up offline planning by incorporating cheap priors through a Schrödinger-bridge diffusion framework. It integrates an image-to-image Schrödinger Bridge (I2SB) with the Diffuser planning approach, introducing three prior types (analytical, learned, random) and evaluating on Maze2D tasks to assess sample and sampling efficiency. The findings show that I2SB can outperform DDPM at very low NFEs due to closed-form sampling and informative priors, but DDPM generally catches up or surpasses I2SB at higher NFEs, with learned priors offering the strongest gains among the priors examined. The work highlights the potential and limitations of prior-guided diffusion for planning, suggesting future exploration of more efficient bridging methods and extensions to higher-dimensional trajectory tasks.

Abstract

Offline planning often struggles with poor sampling efficiency as it tries to learn policies from scratch. Especially with diffusion models, such cold start practices mean that both training and sampling become very expensive. We hypothesize that certain environment constraint priors or cheaply available policies make it unnecessary to learn from scratch, and explore a way to incorporate such priors in the learning process. To achieve that, we borrow a variation of the Schrödinger bridge formulation from the image-to-image setting and apply it to planning tasks. We study the performance on some planning tasks and compare the performance against the DDPM formulation. The code for this work is available at https://github.com/adrshsrvstv/bridge_diffusion_planning.

Planning Using Schrödinger Bridge Diffusion Models

TL;DR

This paper investigates speeding up offline planning by incorporating cheap priors through a Schrödinger-bridge diffusion framework. It integrates an image-to-image Schrödinger Bridge (I2SB) with the Diffuser planning approach, introducing three prior types (analytical, learned, random) and evaluating on Maze2D tasks to assess sample and sampling efficiency. The findings show that I2SB can outperform DDPM at very low NFEs due to closed-form sampling and informative priors, but DDPM generally catches up or surpasses I2SB at higher NFEs, with learned priors offering the strongest gains among the priors examined. The work highlights the potential and limitations of prior-guided diffusion for planning, suggesting future exploration of more efficient bridging methods and extensions to higher-dimensional trajectory tasks.

Abstract

Offline planning often struggles with poor sampling efficiency as it tries to learn policies from scratch. Especially with diffusion models, such cold start practices mean that both training and sampling become very expensive. We hypothesize that certain environment constraint priors or cheaply available policies make it unnecessary to learn from scratch, and explore a way to incorporate such priors in the learning process. To achieve that, we borrow a variation of the Schrödinger bridge formulation from the image-to-image setting and apply it to planning tasks. We study the performance on some planning tasks and compare the performance against the DDPM formulation. The code for this work is available at https://github.com/adrshsrvstv/bridge_diffusion_planning.
Paper Structure (36 sections, 16 equations, 20 figures)

This paper contains 36 sections, 16 equations, 20 figures.

Table of Contents

  1. Introduction
  2. Background & Literature Review
  3. Diffusion models
  4. Diffusion Probabilistic models
  5. Score-Based Generative Modeling through SDEs
  6. Diffusion models in Planning
  7. The Diffuser approach
  8. Architecture
  9. Key Results
  10. Schrödinger Bridges
  11. Schrödinger Bridge Problem
  12. I2SB: Image-to-Image Schrödinger Bridge
  13. Original Schrödinger Bridge SDEs
  14. From SB to I2SB
  15. Implementation
  16. ...and 21 more sections

Figures (20)

  • Figure 1: Diffusion models suffer from expensive sampling and long training times.
  • Figure 2: The DDPM graphical model.
  • Figure 3: Overview of score-based generative modeling through SDEs.
  • Figure 4: Diffuser uses a local receptive field to enforce local consistency while predicting non-autoregressively.
  • Figure 5: Diffuser treats generating a goal-seeking policy as an inpainting problem.
  • ...and 15 more figures