Method
Why this exists, and where every number comes from
Why this exists, where every number comes from, and how accurate it turned out to be when I checked it against the live sources.
Why I built this
I kept running into the same format in different cities. A room of researchers and engineers, a fixed slot in the week, and one rule that makes the whole thing work: you do not just show up, you bring one paper or one repository and explain it to everyone else.
It is a good answer to a real problem. There were 1,166 submissions to cs.RO in the last full month, up 45% on a year earlier. Nobody reads that. Twenty people who each carry one paper properly can cover a useful slice of it every week, and then argue about which part actually matters. The arguing is most of the value.
What I could not find was the index. Every club lives on its own Luma calendar or lab page. Reading lists get published the week of the session and buried the week after. If you move cities, or you want to start a club where there is none, there is nowhere obvious to look. So I built the place to look.
Two things surprised me while assembling it.
- Clubs almost never call papers by their real names. Of 26 reading-list entries I checked, 5 used the arXiv title. The rest were nicknames or descriptions : useful in a room, useless in a search box. That gap is now the main thing this site does: every entry is resolved to its arXiv record, with the club's own wording kept next to it.
- These rooms are invisible until you walk into one. When I first assembled this, Zurich looked empty: world-class labs, a quarterly roundtable, and no findable paper club. Then I went to Zurich, and there it was : a Robot Learning Paper Club running every few weeks, started by someone who was tired of the local robotics scene never being in the same room. It had been running the whole time. I could not find it from a search engine. Tel Aviv and Singapore still read as empty here, and I no longer believe that means they are. If yours is one of the invisible ones, put it on the map.
This is a side project, refreshed twice a week. It is assembled almost entirely from other people's public pages, which is why the full sourcing is below, the datasets are downloadable, and the accuracy failures are printed on this page instead of being quietly corrected. If something about your club or your name is wrong, or you would rather not appear at all, say so and it changes in the next refresh. No justification needed.
Raz Kaplan, gtmascode.dev
The sources
| What | Where it comes from | Records |
|---|---|---|
| Sessions and dates | The clubs' own event pages. Title, date, city and hosts are taken from what the organisers published, and every session on the site links back to the page it came from. | 28 |
| Reading lists | The same pages, read for the pre-reading blocks organisers post before each session. Each entry is then matched to its arXiv record, which is where most of the interesting discrepancies show up. | 39 |
| Paper metadata | The arXiv API. Titles, authors, dates, categories and comments are taken verbatim. Nothing is rewritten. | 2,361 |
| Paper impact | Semantic Scholar, for citation counts, DOIs, and the venue a preprint eventually landed at : 871 papers here carry a conference that the arXiv record alone does not mention. | 2,296 |
| Full text | The paper itself, not a summary of it. 34 parsed from the paper's own PDF with its layout intact; 13 parsed from the paper's own PDF with its layout intact, so tables and equations survive. Each paper's page says which of those it got, so you can discount a weak one. | 47 |
| Slide kits | Each paper's own sections, split on its own headings and quoted verbatim. Nothing is written by a model; the only generated sentence anywhere is a takeaway line, and where that comes from Semantic Scholar rather than the club it says so on the paper's page. | 47 |
| Chapters | Each group's own website or event calendar, read in full and classified by hand. Verified 2026-08-28. | 13 |
| Researchers | Roles as stated by the event host, cross-checked against the public LinkedIn profile for the 11 whose hosts published one. Publication counts come from arXiv author search with co-author clustering, and are left blank where the name does not resolve to one person. Organisers are credited on the chapter they run and are deliberately not listed here. | 38 |
| Trends | One arXiv query per topic per month reading the total-results count. Query strings are printed beside every number. | 408 |
How papers get tagged
Tagging is a published keyword rule per topic, not a model. A paper carries a topic if its title, abstract or arXiv comment contains one of that topic's phrases. This is deliberately auditable and deliberately imperfect: it over-tags papers that mention an area in passing, and it misses work that uses unusual vocabulary. The whole rule set is a single list of phrases per topic. If a tag is wrong, the fix is a one-line change anyone can propose.
The validation gate
Separate from the sampled accuracy run below, every build walks the full papers and repos datasets and quarantines what fails: withdrawn papers, repos that have left GitHub, malformed or duplicate records. Renamed repos are adopted under their new name, and stars are re-read weekly rather than trusted from the last crawl. Nothing is dropped silently; the quarantine list ships with the data.
Accuracy report
Every class of claim is re-fetched from a source independent of the pipeline that produced it (the live arXiv page rather than the API the build used, the live club page rather than our cached copy) and compared. Failures stay on this page.
| Check | Pass | Fail | Skip | Rate |
|---|---|---|---|---|
| Paper titles re-read from the live arXiv page | 12 | 0 | 0 | 100% |
| Paper submission dates re-read from the live arXiv page | 1 | 0 | 0 | 100% |
| Chapter pages still live and still describing a reading group | 13 | 0 | 0 | 100% |
| Upcoming session dates against the live event page | 6 | 0 | 0 | 100% |
| Stated affiliation against the current LinkedIn employer | 7 | 2 | 0 | 77.8% |
| Stored monthly arXiv counts re-queried | 6 | 0 | 0 | 100% |
| The PDF we parsed is the paper we attribute it to | 10 | 0 | 0 | 100% |
| A paper in a resolved author cluster really lists that person | 8 | 0 | 0 | 100% |
| A paper's top recommendation visibly relates to it | 10 | 0 | 0 | 100% |
| Quarantined entries are really gone, and gone from the data | 2 | 0 | 0 | 100% |
What clubs call the papers they read
Separate from pass or fail, because it is not an error on anyone's part. We compared every title a club printed on its event page against the arXiv record for the same paper.
This is the single best argument for an index like this one. A reading list that says “Cross-Embodiment Policy Learning via Representation Alignment” is describing a paper actually called “EMMA: Scaling Mobile Manipulation via Egocentric Human Data”. Both names are useful; only one finds the PDF. Every paper here is resolved to its arXiv record, and the club's own wording is kept alongside it.
| Printed on the event page | The paper it points at | Kind |
|---|---|---|
| PlaNet (2018) | Learning Latent Dynamics for Planning from Pixels | paraphrase |
| Dream to Control (2019) | Dream to Control: Learning Behaviors by Latent Imagination | short name |
| DayDreamer (2022) | DayDreamer: World Models for Physical Robot Learning | short name |
| Director (2022) | Deep Hierarchical Planning from Pixels | paraphrase |
| Dynalang (2023) | Learning to Model the World with Language | paraphrase |
| Learning Dexterous Manipulation from Egocentric Human Videos | EgoMimic: Scaling Imitation Learning via Egocentric Video | short name |
| Scaling Robot Learning with Human Behavior Priors | EgoBridge: Domain Adaptation for Generalizable Imitation from Egocentric Human Data | paraphrase |
| Cross-Embodiment Policy Learning via Representation Alignment | EMMA: Scaling Mobile Manipulation via Egocentric Human Data | paraphrase |
| Human-Robot Co-Design for Scalable Data Collection | Emergence of Human to Robot Transfer in Vision-Language-Action Models | short name |
| World Models from Human Experience | EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data | paraphrase |
| Ego-Exo Transfer: Learning Action from First- and Third-Person Data | DexUMI: Using Human Hand as the Universal Manipulation Interface for Dexterous Manipulation | paraphrase |
| Internet-Scale Human Motion Pretraining for Robotics | EgoVerse: An Egocentric Human Dataset for Robot Learning from Around the World | paraphrase |
| Learning Generalist Robot Policies from Human Demonstrations | DreamDojo: A Generalist Robot World Model from Large-Scale Human Videos | paraphrase |
| A Frame is Worth One Token: Efficient Generative World Modeling with Delta Tokens | An Image is Worth 32 Tokens for Reconstruction and Generation | paraphrase |
Every failure, in full
Not hidden and not silently corrected. Most of these are the interesting part.
| Check | Subject | Site says | Source says |
|---|---|---|---|
| researcher-org | Anthony Zhao | Manycore Tech | Spatial Frontier Club |
| researcher-org | Jie Wang | UPenn | Venture Lab |
Run 2026-08-28. Download the full report, or read the check list above.
What this site does not know
- Clubs that are not public. Plenty of labs run an internal reading group with no page. If yours is public enough to join, add it.
- Citation counts and impact. None are shown, because none are measured here. Paper counts are counts within this corpus and nothing more.
- Complete affiliations. The arXiv API does not return affiliations. Organisation figures come from text mentions and from author blocks in the papers whose full text we hold. Two weak signals, reported separately rather than blended into one confident-looking number.
- Whether a slide kit is right. Kits are extracted sentences. Extraction can pull the wrong sentence. Read the paper.
Corrections and removal
Listings are compiled from pages their owners published. If you run a club or you are listed as a speaker and you want an entry changed or removed, open an issue and it goes in the next build. No justification needed.