A Survey on Diffusion Policy for Robotic Manipulation: a taxonomy and reading list

A reading list, not a library

This repository is a curated collection of resources and papers on diffusion models for robotic manipulation, and the top level listing makes that concrete: it holds a LICENSE, a README and three images, a timeline, a chart of papers over time and a comparison of model architectures. There is no package to install and no code to run. What it offers is a taxonomy. The README presents the work as the first comprehensive analysis of diffusion policies for robotic manipulation, analysing existing methods from three perspectives, data representation, model architecture and diffusion strategy. The framing in the README is that since 2022 diffusion policies have shown consistently superior performance against traditional methods across diverse tasks, and that the field had grown quickly without a survey that synthesised it. Those are the authors' claims, stated in the README, and the list itself is the artefact.

The three axes of the taxonomy

The organising idea is that any diffusion policy can be placed on three independent axes, which is what makes the list navigable rather than chronological. The first axis is data representation, split into 2D representations, 3D representations and heterogeneous data. The second is model architecture, and here the README names three types explicitly: diffusion policies based on a large language model, those based on a small size CNN or transformer, and those based on a VAE or VQ-VAE. The third axis is diffusion strategy, which covers five directions, incorporating reinforcement learning, combining with equivariance, accelerated sampling or denoising strategies, employing classifier or classifier-free guidance, and integration with self-supervised learning. The list of papers is then filed under those headings, which is why the same method can appear to belong to more than one conversation.

What the entries actually give you

Each entry is a paper title with a venue, such as ICLR 2025, ICML 2025, CVPR 2025, ICRA 2025, NeurIPS 2024 or CoRL 2024, and one or more links. The pattern is consistent across the visible entries: a link to the paper on arXiv, sometimes a link to a code repository, and sometimes a project website. Some entries have all three, others only the paper. Among the visible titles are Diffusion Policy Policy Optimization at ICLR 2025 with its own code repository, Latent Action Pretraining from Videos at ICLR 2025, Point Policy at arXiv 2025, SPOT, an SE(3) pose trajectory diffusion for object-centric manipulation at ICRA 2025, RDT-1B, a diffusion foundation model for bimanual manipulation, a modulated transformer diffusion policy model, Reflective Planning, and The Ingredients for Robotic Diffusion Transformers at ICRA 2025. Following the code links from here is the practical route into any of these methods.

Simulators and real hardware are a section of their own

The table of contents ends with a heading for simulation platforms and real world robots, which tells you the list does not stop at papers and code but also points at the platforms used to evaluate these methods. That matters in this field because the comparison between a diffusion policy and an older baseline is only meaningful if both ran in the same simulator or on the same robot, and the survey's framing of future directions depends on what the evaluation infrastructure can support. The README also links a survey paper with a DOI, titled A Survey on Diffusion Policy for Robotic Manipulation: Taxonomy, Analysis, and Future Directions, which is the formal version of the taxonomy the repository is built around.

Images that carry the argument the text cannot

Three images do real work in this repository. The timeline.jpg file gives the history of the field, which supports the claim that the area has been growing since 2022. The number_of_paper_page.jpg chart counts diffusion policy papers over time, and model.jpg compares the architectures of different diffusion policy models, which is a visual version of the architecture taxonomy the text describes. Placing these at the top of the README, above the paper list, is a design choice with a consequence for a reader: the argument for the taxonomy is made visually, and the links below it are the evidence. A reader who wants the reasoning rather than the reading list has to open the survey paper.

An Apache 2.0 list with a July 2026 update

The licence is Apache 2.0, which is permissive enough to republish the list or fold it into internal documentation with attribution, and there is a citation section in the README asking that the survey paper be referenced. The repository is not archived and the last push was on 2026-07-13, so entries are still being added. There are no GitHub releases, which is expected for a repository whose only artefact is a document, and the primary language field is empty because there is no source code. The obvious limitation is that the value of the list depends entirely on its coverage and its curation, and neither can be checked without reading it: there is no count of papers, no search function, no tags, and no indication of which entries the authors consider foundational rather than merely recent. The visible entries are weighted towards 2024 and 2025 work.

Editorial conclusion

The repository suits a researcher or engineer entering diffusion policies for manipulation who needs a map of the field and a route into the code for individual methods, since entries link out to papers, repositories and project pages. It does not suit someone looking for a runnable implementation, a benchmark harness or a comparison of results, because it holds only a README and three images. Before citing it, follow the survey paper link the README points to, and check that a method you care about appears under the architecture family you expect, since the taxonomy is the authors' and categories overlap in practice.

Frequently asked questions

Official sources

1. iLearn-Lab/VLA-Diffusion-Policy-Robotics on GitHub
2. Issues
3. License: Apache-2.0
4. README