Guide

How to start a robotics paper reading club

If your city has robotics labs and no reading club, the missing ingredient is almost never interest. It is one person willing to pick a date and present first. Here is what the 11 groups in this index do, and what they had to fix.

The five decisions that define your club

1. A fixed slot, published in advance

Every durable group in this index has a slot you can predict without checking: Saturday afternoons in San Francisco, Friday noon at Mila, Monday evenings in Berlin. A moving date reads as an optional date. Publish the next one before the current one ends.

2. A filter on what counts as a paper

This is the decision that keeps a robotics club about robotics. The Montreal group requires a real or simulated robot deployment. The UT Austin group scopes each term to a theme and pulls from CoRL, RSS and ICLR. The MIT LINGO lab votes a theme per round and spends several weeks inside it, including textbook chapters, which is the right call when your group needs shared foundations more than novelty.

3. A mix of people, not one lab talking to itself

The Zurich club exists because robot learning there was scattered across academia, big tech and startups that rarely shared a room. That is the thing worth engineering for. A session where a second-year PhD student, an engineer shipping arms to factories and someone switching in from computer vision all read the same paper produces objections none of them would have reached alone, and it is the reason people come back.

Practically: make attending easy and presenting easy, and do not gate the door. Say on the page that anyone can come and listen. Where spots are limited, giving priority to people who have presented before rewards contribution without turning it into a requirement.

4. A structure the presenter has to fill

Unstructured talks drift into a slide-by-slide walk through the paper's figures. Give presenters a skeleton: the problem, what came before, the method, what was measured, where it breaks, and the one line worth remembering. Utah's guideline, say the assumptions out loud before showing results, is the single highest-yield rule we found, because assumptions are where robotics papers usually fail and where they are usually quietest.

5. A public record

Publish the reading list before the session and the takeaways after. The San Francisco club posts numbered pre-readings with a one-line summary of each, then a written recap afterwards. That record is what lets someone join at session 20 without feeling behind, and it is what makes your club findable at all.

Week one, concretely

1. Pick the slot and the next four dates. Put all four on the page.
2. Write the filter in one sentence and put it above the fold.
3. Choose three papers yourself for session one. Do not crowdsource the first list.
4. Present the first one yourself, badly, in ten minutes. This sets the bar low enough that others volunteer.
5. Before people leave, name the presenter for session two out loud.

What kills new clubs

• No named presenter for the next session. The most common failure. Leave the room with a name.
• Scope creep into general AI. Robotics people leave when three sessions in a row have no robot in them.
• One organiser. Every long-running group here has at least two people whose names are on the events.
• Sessions that run long. Forty-five minutes of talk needs a hard stop; discussion is what people came for.
• No archive. A group with no public record cannot recruit, and cannot prove it exists.

Questions people ask before starting one

How often should a robotics reading club meet?

Weekly works when there is a dense local scene and a rotating pool of presenters; the San Francisco club has run 33 weekly Saturday sessions. Fortnightly is the safer default for a university group: UT Austin runs one paper every two weeks with a 45-minute presentation and 15 minutes of questions. Monthly tends to lose the thread, because members forget what the group concluded last time.

How many people should be in the room?

Eight to twenty for a discussion group. Below eight, a single absent presenter cancels the session. Above about thirty the format changes: the Zurich Roundtable caps at 250 and is explicitly a panel and networking night, not a paper discussion. If you want both, run them as separate events.

Do people have to present to come?

At the clubs in this index, no. One presenter plus discussion is the standard lab-seminar shape, and the open clubs run a keynote or a couple of talks with an open floor after. Some cap the room and give priority to people who have presented before, which rewards contributing without making it a condition of entry. Requiring everyone to present would cost you exactly the people worth having: the ones who came to find out whether this field is for them.

How do you get more than one lab in the room?

Say on the page who it is for, and make it plural. The Zurich club was started because academia, big tech and startups in the city were all pushing robot learning forward and rarely shared a room. Alternating who presents between a university group and an industry team does more for the discussion than any rule about attendance, because the objections come from different places.

How do you choose which papers to read?

Set a filter and publish it. Montreal requires every presented paper to include a real or simulated robot deployment. Utah requires the presenter to state the paper’s three to five core assumptions before showing any results. A filter is what stops a robotics club drifting into a general machine-learning club by month four.

Do you need a venue and a budget?

A room and a screen. The clubs in this index run out of university auditoriums, startup offices, campus buildings and co-working spaces. Food is the one expense worth having, because it converts a 45-minute seminar into the two hours where people actually talk to each other.

When it is running

Add it to this index and it appears on the chapters map with your cadence, format and link. There is no fee and no approval queue beyond a check that the page is real.