ICLR-2025

Things that worked:

• The workshop being not on the last day. We knew this and successfully changed ICLR’s original decision to put it on the last day, but just to emphasize: the overall venue attendance on that day was way lower.

• Grouping virtual talks as much as possible. We had two virtual invited talks one after another, and all video orals in one block. The more they are grouped, the less switching overhead and the less the chance of equipment issues.

• Putting high-profile talks after the breaks and at the end of the day. We started the day with the Best Paper Award talk, had Chelsea’s talk right after lunch, and Coline’s talk about Gemini Robotics at the end, encouraging people to be back on time and stay until the end.

• Robot demos, especially quirky ones. Even though we had only two exhibitors, they did a great job spicing up the workshop. If the exhibitors are companies, they are incentivized to make their exhibits exciting.

• Theming around a broad phenomenon (human-level abilities) rather than a class of techniques/approaches (e.g., RL or pretraining) has likely helped. Certainly, the variety of the papers was very broad.

• Keeping close track of time. We finished almost every workshop section on time, and overran the scheduled time by only ~10 min.

• Having remote speakers join the Zoom session way in advance of their talk (20 min). Davide had issues with his slides that took a while to resolve, but his session was still on time thanks to his joining early.

• Putting up the workshop schedule early (at least 10 days in advance). We did this, but the Gen AI for Robotics workshop didn’t until shortly before the workshop day. They ended up having some technical issues with schedule visibility that they didn’t have time to fix. As a result, nobody could find their schedule even on the workshop day itself and thus had no idea what the content would be. This seems to have affected that workshop’s attendance badly.

• Testing all the demo equipment (the evening) before the workshop. We did this, and it revealed issues that would be at best very difficult and stressful to address on the day of.

• Insisting on speakers’ in-person participation. Besides simplifying the workshop logistics, the engagement levels for live talks and poster session tended to be noticeably higher than during the virtual ones. 1-2 remote invited talks out of 7 is ok, but no more than that.

• Alternating content type in adjacent workshop sections was likely helpful. E.g., we had a contributed orals sessions followed by a poster + robot demo session followed by invited talks followed by lunch followed by more invited talks followed by another poster + demo session followed by a panel and a final round of invited talks. This kept the workshop from becoming monotonous. We ended up having a full room throughout the day.

Things to be aware of/improve:

• Local demo equipment requirements and costs. E.g., one of the demos had high requirements for the internet speed, which weren’t very clear at the beginning and costly to address at the end.

• Robot exhibitors bailing on us. In our case, two bailed within a week of the workshop. Unclear how this can be prevented, but we should be prepared for this and be sure to have enough exhibitors to begin with to tolerate the last-minute losses.

• Unresponsive invited speakers. Once a speaker has agreed to give a talk, we shouldn’t be shy to relentlessly bug them about confirming their talk time, clearly communicating the deadline for doing this (e.g., no closer than 3 days before the workshop) to give ourselves a chance to find a replacement if needed.

• Soliciting sponsors early. We did this mostly late (1 month in advance), and although we secured enough money in the end, this could have ended worse than it did.

• Assigning responsibilities to organizers in advance, e.g., sponsorship, webpage maintenance, equipment rental/issues, etc. We did this for some organizers but not others. As a result, some of us ended up ad-hoc on point for too many things, while others ended up underutilized.

• Invited talk duration. We had 34 min for each and suggested 30 min for the talk + 4 min for questions. 30 min for the talk was too much for most of them; 25-26 min would have been enough. 4 min for questions was great though. So, we should aim for 30 min total in the future.