diffusion-literature-for-robotics is a reading list, not a library

Two files, one of which is the entire contribution

The repository contains exactly two entries at the top level, LICENSE.md and README.md. There is no code, no configuration, no build system and no data.

The README describes what it is in one sentence: a brief summary of essential papers and blogs on diffusion models, alongside a categorized collection of robotics diffusion papers and useful code repositories for starting your own diffusion robotics project.

That makes it a reading list, and it is worth being clear about that before anyone adopts it. Adopting a reading list means reading it, bookmarking it and checking it later. There is no dependency to add, no version to pin and nothing to upgrade. The MIT licence on the repository covers the text of the list itself, which is why the licence file is one of only two files present.

The practical value is in the organisation. The topics attached to the repository are diffusion-models, grasping, imitation-learning, motion-planning, offline-rl, reinforcement-learning, robotics, score-based-generative-modeling, summary and video-diffusion-model, which is a fair description of what the list covers.

The reading order starts with blogs, then the five theory papers

The first section, titled Learning about Diffusion models, is explicitly a beginner path. The README says that while many tutorials exist, it collects some of the best introduction blog posts and videos.

Five entries are listed. Two are videos: an introduction video called What are Diffusion Models?, which introduces the general idea and some high-level math, and a paper explanation video called Diffusion Models | Paper Explanation | Math Explained, which covers the math and notation in detail with visual aid. Three are written: Yang Song's score-based blog post, described as coming from one of the most influential authors in the area, Lilian Weng's in-depth post on the theory with a summary of how the model improved over time, and Understanding Diffusion Models, an in-depth paper explaining the model from both perspectives with detailed derivations.

For readers who would rather start with the originals, the README then lists the most important diffusion theory papers: Sohl-Dickstein and colleagues on deep unsupervised learning using nonequilibrium thermodynamics from ICML 2015, Ho and colleagues on denoising diffusion probabilistic models from NeurIPS 2020, Song and colleagues on score-based generative modeling through stochastic differential equations from ICLR 2020, Ho and Salimans on classifier-free diffusion guidance, and Karras and colleagues on elucidating the design space of diffusion-based generative models from NeurIPS 2022.

The README also points to a general list of all published diffusion papers at the Whats the score? site, which is the honest place to look for completeness.

Nine categories map how the robotics literature subdivides

The second half of the README is the robotics part, and it opens with an honest admission: since modern diffusion models have been around for only three years, the literature about diffusion models in robotics is still small, but growing rapidly. The list covers papers published at conferences or uploaded to arXiv.

The table of contents divides them into nine areas. Imitation Learning and Policy Learning comes first, followed by Video Diffusion in Robotics, then Online RL, Offline RL and Inverse RL. After that come World Models, Task and Motion Planning, Tactile Sensing & Pose Estimation, and Robot Design and Development. A separate section collects Code Implementations, and a final section is titled Diffusion History.

That structure is the most transferable part of the repository. A category list is a claim about how a field is organised, and it is a better guide than a flat alphabetical pile of papers. If you are looking for work on grasp pose estimation under tactile sensing, the fact that the list has a category for it tells you that enough work exists to be worth a category.

The README does not annotate the entries with claims about which approach works better, and the sections as visible in the available text are lists of citations rather than comparisons. Any synthesis beyond the grouping is on the reader.

The imitation learning section is the densest and the most dated

Imitation Learning and Policy Learning is where the README spends most of its visible length, and the entries are recognisable names in diffusion policy work.

Among them are 3D Diffuser Actor for policy diffusion with 3D scene representations, One Model to Drift Them All for physics-informed conditional diffusion in driving, Sparse Diffusion Policy, Equivariant Diffusion Policy, GenDP for category-level generalisable diffusion policy, EquiBot for SIM(3)-equivariant diffusion policy, DiffuseLoco for real-time legged locomotion control from offline datasets, MaIL for improving imitation learning with Mamba, Generative Image as Action Models, Variational Distillation of Diffusion Policies into Mixture of Experts, Language-Guided Manipulation with Diffusion Policies and Constrained Inpainting, and TEDi Policy for temporally entangled diffusion for robotic control.

Two observations. The first is concentration: a large share of the entries are from the 8th Annual Conference on Robot Learning and from 2024 arXiv preprints, including one 2024 preprint listed under two different entries in the visible text, which suggests the list was compiled by pasting in items as they were found rather than by a strict deduplication pass.

The second is what that implies. These entries are 2024 work in a repository whose last push was on 2026-09-22. Two years of diffusion policy papers are not visible in the text available here, which is exactly the caveat a reader has to check before treating the list as a survey.

Code bases and diffusion history are pointers, not implementations

The README's table of contents ends with two sections, Code Implementations and Diffusion History, and their role is different from the rest of the document.

Code Implementations is where the repository points to code repositories for starting your own diffusion robotics project. The description of the project as a whole calls these useful code repositories for starting your own diffusion robotics project, which is the clearest statement of intent in the README: the author's aim is to shorten the distance between a newcomer and a first working implementation.

Diffusion History is a section for the origins of the field, which is where the foundational papers listed earlier connect, since Sohl-Dickstein in 2015 and the score-based framing of Song and colleagues in 2020 are the papers that make the timeline readable.

The quote the README opens with comes from the same lineage. It attributes to Yang Song, in Score-Based Generative Modeling through Stochastic Differential Equations, the line that creating noise from data is easy while creating data from noise is generative modeling. The link points at the arXiv PDF for that paper.

Neither of these two sections is visible in full in the text available here, so the concrete repositories and the timeline entries themselves have to be read in the repository rather than inferred from the headings.

No releases, MIT licence, and a branch that keeps moving

The repository has no GitHub releases, which is unremarkable for a document like this and means there is nothing to version or pin. The repository is not archived, and the last push was on 2026-09-22.

That date creates the interesting tension in this project. A branch pushed in September 2026 suggests the file is still being edited, yet the visible paper entries are 2024 conference proceedings and arXiv preprints. Both things can be true at once: the readme structure may be maintained while the paper lists lag, or the visible section may simply be the part that was not part of this extract.

There is no changelog, no per-category count, no indication of how many papers are listed in total, and no statement about which venues are covered beyond conferences and arXiv. The README does mention that a general list of all published diffusion papers exists elsewhere, which is an implicit acknowledgement that this list is not complete.

The licence field on the repository says MIT and the LICENSE.md file is present, so the text can be reused and adapted with attribution, which is convenient if you want to fork the list into a reading group document or a lab wiki.

Editorial conclusion

diffusion-literature-for-robotics suits a researcher entering diffusion models for robotics who wants one curated entry point instead of searching arXiv, and the category structure is a good map of how the field subdivides. It does not suit anyone who needs an implementation, a benchmark or a maintained survey, because the repository contains only a README and a licence file. Before relying on it, check the publication years of the papers it lists against the date of the last push, because the visible entries are 2024 conference and arXiv work while the branch was pushed on 2026-09-22, and treat the list as a starting point to verify rather than a complete survey.

Frequently asked questions

Official sources

1. Issues
2. License: MIT
3. mbreuss/diffusion-literature-for-robotics on GitHub
4. README