iLearn-Lab/VLA-Diffusion-Policy-Robotics

What it solves

Researchers in robotic manipulation lack a single, organized reference that maps the fast-growing field of diffusion policies. Since 2022, diffusion-based methods have consistently outperformed traditional approaches on many robot tasks, but the papers are scattered across venues and use inconsistent terminology. This project fills that gap by offering the first comprehensive survey that systematically categorizes and summarizes the existing literature.

How it works

The survey examines diffusion policies from three angles: how data is represented (2D, 3D, or heterogeneous), how the model is built (large language model based, small CNN/transformer based, or VAE/VQ-VAE based), and which diffusion strategy is used (e.g., incorporating reinforcement learning, equivariance, accelerated sampling, classifier guidance, or self-supervised learning). It then lists relevant papers under each category, providing links to the papers, code, and project pages. The project also includes a timeline and figures showing the growth of publications and comparisons of model architectures.

Who it’s for

This is aimed at researchers, graduate students, and practitioners in robotics and machine learning who want to quickly understand the state of diffusion-based robot manipulation, find relevant papers by category, or identify research gaps and future directions.

Highlights

• First comprehensive analysis of diffusion policies for robotic manipulation, covering data representation, model architecture, and diffusion strategy.
• Extensive curated list of papers with direct links to arXiv, code repositories, and project websites.
• Organized taxonomy with three architecture families: LLM-based, small CNN/transformer based, and VAE/VQ-VAE based diffusion policies.
• Includes a timeline and statistical figures visualizing the growth and architectural comparisons of diffusion policy methods.
• Companion survey paper available with deeper analysis and future directions.