This repository provides a comprehensive survey of diffusion policies for robotic manipulation, categorizing existing methods by data representation, model architecture, and diffusion strategy. It serves as a valuable resource for researchers and practitioners in robotics and machine learning seeking to understand and apply diffusion models for complex manipulation tasks.

How It Works

The survey systematically analyzes diffusion policy research, classifying methods into categories such as Large Language Model (LLM) based, small CNN/Transformer based, and VAE/VQ-VAE based architectures. It also details diffusion strategies like reinforcement learning integration, equivariance, accelerated sampling, and classifier-free guidance. This structured approach offers a clear taxonomy for understanding the landscape of diffusion policies in robotics.

Quick Start & Requirements

This repository is a survey and does not have a direct installation or execution command. It links to numerous research papers and their associated code repositories, which will have their own specific setup instructions and dependencies (e.g., Python, PyTorch, specific hardware like GPUs).

Highlighted Details

• Comprehensive taxonomy of diffusion policies for robotic manipulation.
• Analysis across data representation (2D, 3D, heterogeneous), model architecture, and diffusion strategy.
• Extensive list of recent papers (2022-2025) with links to code and websites.
• Detailed tables of simulation platforms and real-world robots used in diffusion policy research.

Maintenance & Community

The primary contribution is a survey paper, with authors affiliated with HITSZ-Robotics. Links to related code repositories are provided for individual papers.

Licensing & Compatibility

The survey itself is presented as a research document. The licensing of the linked code repositories varies per project and must be checked individually.

Limitations & Caveats

As a survey, this repository does not provide a unified codebase or framework. Users must refer to individual linked papers for implementation details, dependencies, and potential limitations of specific diffusion policy methods. The rapid pace of research means some cited papers may still be in pre-print or under review.