Collaborative_Perception is a curated paper list for V2X perception

A reading list with two jump targets

The repository does one thing and says so. It is a paper digest of recent advances in collaborative, cooperative and multi-agent perception for V2I, V2V and V2X autonomous driving scenarios, and the papers are listed in alphabetical order of the first character.

The structure is visible in the jump links at the top, and there are only two: Method and Framework, and Dataset and Simulator. Everything else hangs off the recommendation section, which is organised by resource type. There are Position entries, which are surveys and framing papers. There are Talk entries, which are recorded conference talks and tutorials. There are Library entries, which are codebases. There are Workshop entries. And there are Background entries for the reading you need before any of it makes sense.

The repository itself is three files, `README.md`, `mode.png` and `direction.png`, so the images are not decoration. They illustrate the two taxonomy diagrams the README refers to, the typical collaboration modes and the possible optimisation directions, and they are how the author organises a literature that otherwise has no agreed structure.

The primary language and the licence are both recorded as unknown, which is accurate for a document of this kind. There are no releases, the default branch is main, and the last push was on 2026-10-08.

The benchmark tables were removed on purpose

The most valuable paragraph in this README is the apology for what is no longer there.

The author writes that they find it hard to fairly compare all methods on each benchmark, since some published results are obtained without specified training and testing settings, or even with modified model architectures. In fact, many works evaluate all baselines under their own settings and report them. Therefore, it is probably to find inconsistency between papers. Hence, the collection and reproduction of all the benchmarks in a previous update was discarded.

That is a methodological decision with real consequences, and it is worth understanding before you rely on the list. A table of benchmark numbers assembled by one person from many papers looks like a leaderboard and behaves like one, but the numbers may not have been produced under the same conditions. Removing the table removes a false comparison, and the cost is that a newcomer has no reference table at all.

The README does not pretend the data is gone. It points to an archived version of the repository at a specific commit, 1be25908aea0a9f635ff4852b3a90729cf2b6aac, and says that if you are interested you can find a bunch of results there. So the historical tables exist, clearly labelled as belonging to a previous state of the document, which is the honest way to keep them around.

The libraries section is the most immediately usable part

If you are going to do something rather than read something, the Library entries are the section to start from, and each one links both code and documentation.

V2Xverse is a codebase for V2X-based collaborative end-to-end autonomous driving. HEAL is an extensible framework for open heterogeneous collaborative perception. OpenCOOD is an open cooperative detection framework for autonomous driving. CoPerception is an SDK for collaborative perception. OpenCDA is a simulation tool integrated with prototype cooperative driving automation.

Five projects, and they cover different needs. Two are frameworks you extend, HEAL and CoPerception. Two are toolchains for training and evaluation, V2Xverse and OpenCOOD. One is a simulation environment, OpenCDA, which is the piece you need before any of the others can be evaluated on anything.

The heterogeneity in the title of HEAL is the thread running through the field. A roadside unit, a vehicle and a drone do not see the same thing from the same place, and the coordination problem is not only about what to send but about how to reconcile observations that disagree. That is why the library list matters as much as the paper list, since the frameworks are where that reconciliation is implemented.

Position papers first, then talks, then background

The recommendation section leads with Position entries, and the titles tell you the shape of the argument being had in the field. Multi-Agent Embodied Autonomous Driving: From V2X Information Exchange to Shared World Models. Infrastructure-Centric World Models, about roadside perception rather than vehicle-centric. When Autonomous Vehicle Meets V2X Cooperative Perception: How Far Are We?, which is the calibration paper. Research Challenges and Progress in the End-to-End V2X Cooperative Autonomous Driving Competition. Automated Vehicles Should be Connected with Natural Language, which is the odd one out and the most interesting. Collaborative Perception Datasets for Autonomous Driving: A Review, and a long tail of surveys including V2X Cooperative Perception for Autonomous Driving: Recent Advances and Challenges and A Survey and Framework of Cooperative Perception: From Heterogeneous Singleton to Hierarchical Cooperation.

The Talk entries are conference talks and tutorials rather than papers, and several are worth more than a survey for orientation. The Waabi V2V Communication tutorials at CVPR 23 and CVPR 24 are there, along with The Ultimate Solution for L4 Autonomous Driving, When Vision Transformers Meet Cooperative Perception, a talk on collaborative perception under communication interruption, and Collaborative and Adversarial 3D Perception for Autonomous Driving.

The Background section is the one to read first if you are new. It holds a review of point cloud object detection in intelligent agents, a review of 3D object detection for autonomous driving, DACOM on learning delay-aware communication for multi-agent reinforcement learning, and a survey of multi-agent reinforcement learning with communication. That is the vocabulary of the field in four links, and it is the shortest path into the rest of the list.

Workshops, and the method list that pairs papers with code

The Workshop entries point at community events rather than papers: Co-Intelligence at ECCV 24, CoPerception at ICRA 23, and ScalableAD at ICRA 23. For a field this young, the workshops are often where the unpublished work is, and where the people who will write the next survey are.

The Method and Framework section is the long tail, and it follows a different rule. Each entry is a method with a paper link and, usually, a code link. ACCO asks whether discretization fusion is all you need for collaborative perception. AR2VP is dynamic V2X autonomous perception from road-to-vehicle vision. BOLT is online lightweight adaptation for preparation-free heterogeneous cooperative perception. CPPC is a compact message unit for communication-efficient collaborative perception, and it is the one entry that links a paper and a review on OpenReview with its code marked as struck through, which is a useful signal that the code is not available. CP-FREEZER is a paper on latency attacks against vehicular cooperative perception, and CMP is cooperative motion prediction.

The section carries one annotation convention, and it matters when you read the list. The note says that {Related} denotes that it is not a pure collaborative perception paper but has related content, so those entries are pointers to adjacent work rather than core results.

And the entries are ordered, like the rest of the digest, by the first character of the name.

What a paper list can and cannot do for you

The value here is orientation, and the limits follow from what the repository is.

It is not a survey with an argument, and it does not summarise the papers. An entry gives you a title, an arXiv number and a code link, and nothing else, so the reading is still yours to do. It is also not exhaustive, and the selection reflects one person's judgement about what matters. And with the benchmark tables removed, it cannot tell you which method is best, only which methods exist and where the code is.

There is also a licensing question to settle before you build on it. The repository records no licence and no licence file is present, so the default under copyright applies, which means you can read it and link to it, and you should not assume you may republish its text as your own list. For a personal reading list that is not a problem. For a company that wants an internal curated list of the same papers, the honest options are to write your own list with the same links or to ask the author.

The alternative approach is to do the reading from the venues directly, following the arXiv listings and the conference programs for the workshops named here. The difference is exactly the work the digest saves you: this list has already grouped the field by resource type, separated surveys from methods, and, most usefully, pointed you at the five codebases that let you reproduce the ideas rather than only read about them.

Editorial conclusion

Adopt this repository when you are entering V2X collaborative perception and need a map of the field in one page, especially when you want to know which open source toolkits exist alongside the papers. Do not treat it as a benchmark or a comparison source, because the author removed the benchmark tables, and do not copy its content into your own work without checking the licence, which the repository does not state. Verify first that the papers it lists are still the ones you need, given that the last push was on 2026-10-08, that the code repositories it points at are the maintained forks, and that the archived version is where the numbers you remember actually live.

Frequently asked questions

Official sources

1. Issues
2. Little-Podi/Collaborative_Perception on GitHub
3. README