The 2019 International Conference on Robotics and Automation (ICRA)

A photo I took of the ICRA 2019 conference banquet. Look carefully at what the performers are doing.

Last week, I attended part of the 2019 International Conference on Robotics and Automation (ICRA)¹, held in Montréal, Canada at the Palais des congrès de Montréal. I also attended ICRA 2018 last year, in Brisbane, Australia. I have blog posts which summarize my experience at ICRA 2018: the day before, day 1, day 2, day 3, day 4, and day 5. This time, I am not going to write my excessively detailed posts. First, more posts would be somewhat redundant. Second, ICRA is normally five days: workshops and tutorials are two days, while the other three are the “main” part of the conference with plenary/keynote talks and poster sessions. I was only at ICRA 2019 for the main part (May 20-22), and I prefer to write detailed summaries for when I attend a conference in its entirety. Finally, I have deadlines coming up, and my collaborators would be furious if they saw that I was spending valuable time pontificating the intricacies of ICRA rather than doing whatever they need me to do.

Of course, this post is probably on the long side anyway, but don’t tell anyone. In my defense, the majority of it was written on the return flight from the conference, and being in cramped coach class means it is nearly impossible to get real work done. Thus, this post will be structured in a similar style as my IJCAI 2018 post, except perhaps with better organization.

I was not even planning on attending ICRA 2019. I have a paper there, but my colleague Xinlei Pan at UC Berkeley is the first author and was going to present. Unfortunately, he got hit with “visa issues,” and was unable to leave the United States. As someone who happily calls himself a globalist,² this is disappointing, and I will comment on this in future blog posts.

As is my tradition for traveling, I arrived at the San Francisco International Airport early so that I could enjoy a nice dinner, followed by a long time block which allows me to read books (this is how I read so many) or assiduously work on blog posts. As evening struck that night, I boarded my red-eye Air Canada flight to Montréal.

In Montréal

Once I arrived on May 19 (the day before ICRA 2019 started), I took an Uber to the conference venue. I was there to meet the two assigned sign language interpreters, who I had been emailing and texting beforehand. I followed my pre-conference logistics checklist, and we were there to meet the venue’s employees to figure out the best seating position and the lighting situation in the lecture rooms.

Due to Xinlei’s visa situation, I only had 1.5 month’s notice for attending ICRA. Thus, Berkeley’s DSP and I had to scramble to get the logistics organized. Fortunately, DSP has been getting better and better, and was able to obtain the sign language interpreters and contact the conference organizers to get identification for them.

It was a far cry from what happened at UAI 2017. Read this blog post if you want to know the details.

The interpreters knew the American version of sign language. Since they were from Canadian cities, they were additionally able to teach me the signs for Montréal, Toronto, and Quebec. My sign language education is thus augmented by attending international conferences.

While we obtained conference badges, the lighting situation in the lecture rooms was less than ideal. It was too dark. We spoke with the relevant person in charge, and she said there was no way to fix a particular light towards a location (i.e., where the sign language interpreter would be located). I’m surprised the venue lacked that capability. At IJCAI 2018, it was easy for the venue employees to adjust the lights. I’m not sure if the Brisbane Convention Centre lecture hall has that capability, but it was brighter there so we did not have to worry.

Once I had finished discussing logistics with the sign language interpreters, I headed towards my hotel room. Or, actually, my AirBnB. I was lucky to join an AirBnB with several Berkeley students, but was unlucky with what happened two days before I arrived to Montréal: we got canceled. We were re-assigned to two different AirBnBs much farther away from the conference venue. At least the route (about a 30 minute walk) was almost entirely on one street, and seemed safe. To be honest, maybe I will just stick with hotel rooms in the future? I miss having fancy breakfasts and fitness centers, like the hotel in Sydney, Australia for UAI. That one was easily the best hotel I’ve been to for a conference, outclassing the ones from Brisbane and Stockholm.

For dinner, I got recruited by a Berkeley postdoc to eat at a famous place serving poutine. I did not know what that term meant, and upon arriving there, I realized that poutine is a dish consisting of french fries, cheese, brown gravy, and some toppings. I ordered one with grilled onions, and was able to eat all the onions. But, try as I might, I just could not eat even a fraction of the french fries. This is a popular dish associated with Quebec cuisine, but wow, how do people manage to eat this stuff?

I appreciate getting to eat poutine once. That night was my first — and last — time eating poutine.

After “dinner,” I went to a grocery store that I had the tremendous fortune to pass by. As luck would have it, there were fresh blueberries and baby carrots in stock. I bought some to eat back at the AirBnB.

I don’t think I can ever remember feeling happier to consume blueberries and baby carrots.

Day 1

On the first day of the conference, we had our opening ceremony. (This was different from Brisbane last year, when we had tutorials on the first day, followed by the opening ceremony in the evening.) There were two sets of performers: female dancers, and male singers. The dancers seemed very intense, and were impressively synchronized throughout their performance. One of the sign language interpreters commented that the applause was stronger for the male singers, who conducted a less physically strenuous task. Was there some sexism here?

After the performances, Greg Dudek gave the opening remarks. As usual, he commented on the large number of papers accepted and the growing size of the conference. Also, as usual, there was some not-so-subtle patriotism. Dudek presented the number of conference papers with respect to a country’s population, which meant Canada was in first place and ahead of the United States. Then we realized that he was half right. Canada is ahead of the United States, but not first. On the next slide, he corrected himself by showing Switzerland and Singapore in first and second place, respectively. It looks like Switzerland won the prize of papers with respect to a country’s population. Last year, Singapore won it. What happened, Singapore???

Then we had the plenary talks, followed by the keynotes. In ICRA terminology, plenaries are arguably more prestigious than keynotes, because there is exactly one plenary talk in the morning for each of the three main conference days, and this is followed by two keynotes held simultaneously.

That day, Yoshua Bengio gave the plenary talk. Oh yeah. Actually, there seem to be a lot of similarities with this year’s plenaries and last year’s:

• One famous white man: Rodney Brooks last time, Yoshua Bengio this time. Both are eminent leaders in the academic community.
• One famous man of Indian descent: Mandyam Srinivasan of Queensland last time, and Vijay Kumar of the University of Pennsylvania. Both are known for biological or aerial robots, in some sense.
• One famous white woman: Raia Hadsell of DeepMind last time, and Raquel Urtasun of Toronto/Uber. Both are famous Deep Learning researchers, and both talked about self-driving cars and navigation.
• Two out of three plenary speakers were from the conference venue’s country in some sense, either via citizenship or by working there as a permanent resident. Brooks and Srinivasan are associated with Australia, and Bengio and Urtasun are associated with Canada.

Folks, I think I see the pattern here.

Bengio’s talk was surprisingly short, and I think the reason became clear when Dudek said that Bengio had to catch a flight after his talk. I think the audience still enjoyed it, and Dudek said we could tell future generations of researchers that “we got to see Yoshua Bengio give a talk.” I don’t think that comment applied to me, because this was the second talk of his I attended; he gave one in Fall 2015 at UC Berkeley, recruited by Pieter Abbeel. That was a standing room only talk, though surprisingly, I think I remember more from that talk than from his plenary.

After Bengio’s plenary, we had poster sessions and coffee breaks. Sadly, there were no free lattes, unlike in Brisbane. Darn. In terms of the food, the coffee breaks followed a pattern of serving two main dishes: one pastry-based and one fruit-based.

ICRA 2019’s poster sessions followed the same pod-based system as last year. From browsing the topics, it was clear that deep learning, imitation learning, reinforcement learning, and autonomous driving were hot topics. Someone with more knowledge of the paper and keyword statistics might correct me on this, but my impression is that there was a similar distribution of research topics as last year.

The first set of keynote talks occurred in the afternoon. I attended the one based on industry and venture capitalism. Annoyingly, this room (not the same as the room for Bengio’s plenary) was also dark, and I did not get much benefit out of the talk as it was hard to understand the sign language interpreters. The speaker concluded with a slide that recommended six books to read. I read two out of the six: The Hard Things About Hard Things and The Lean Startup. You can see my summaries here.

After that, we had more poster sessions, and then the welcome reception in the same room. Interestingly, they limited us to two alcoholic beverages each, unlike in Brisbane where we had unlimited servings. That seems fair, as we don’t want people to overdo it. The food was also reasonable, with some intriguing smoked meat dishes and (ugh) poutine. It was cocktail-style, so I walked around sampling the food (and dessert, such as iced maple syrup) and chatting with people who I knew. Eventually, I found Ken Goldberg there with other Berkeley students, so I was able to join their group, and we took a few group photos.

After the welcome reception stopped serving food, I went back to our AirBnB. I needed to stay up late to finish some grading for CS 182/282A.³

Day 2

I woke up early to go for a run in a park near the conference venue. The trail was nice, but with respect to running routes, I give the edge to Brisbane. There was a trail underneath the highways that gave a great view of the Brisbane coastline.

I arrived to the conference venue in time for the second plenary talk, by Vijay Kumar. I was fascinated by his videos showing aerial and flying robots. And, as is typical nowadays for many robotics keynotes, Kumar talked about industry investment in robotics and AI. I left his talk thinking that I needed to (a) read some of his papers, and (b) know more about robotics-related business.

The keynotes directly followed the plenary. I chose to attend CMU Professor Matt Mason’s talk about failures in research. He was Ken Goldberg’s PhD advisor, and to the surprise of no one, Ken was sitting in the audience and taking pictures. Matt Mason was also a PhD student in the same group at MIT as John Canny was back in the 1980s. Mason and Canny were advised by the great Tomás Lozano-Pérez.⁴ It’s odd that Mason was talking about failures, because he’s had an enormously successful career. I guess that was the point!

Incidentally, it was held in the same room as the keynote from yesterday. There was more lighting near the center of the room, so the sign language interpreters and I figured out a way to position ourselves so that I could at least understand a fraction of the talk’s contents. Before the talk, we were amused when an observer told us that one of the interpreters (who was facing me) was seated in the wrong direction.

Later, for one of the poster sessions that day, I helped other members of Ken Goldberg’s lab to present the Fog Robotics paper. The metaphor of “fog” is that it’s an intermediary between the “cloud” and the “edge,” but to really understand it, read the paper. Note that Ken Goldberg is well known for cloud robotics research, and published a well-regarded survey on the topic. My colleague Ajay Tanwani was supposed to present since he is the first author, but as with Xinlei, he got snagged up with visa issues. Sigh, what in the world is going on with visa issues?

While helping to present the paper, I met someone who studies gesture recognition. She was mesmerized by the sign language interpreting, and asked to take a video of us talking (we gave permission). Meanwhile, I enjoyed describing Fog Robotics, and hope to be more involved in that research over the summer.

After the poster sessions, we had the conference banquet! Unlike in Brisbane, the conference banquet was a sit-down dinner; see the picture at the top of this post. Normally, I get into panic mode, because the most likely outcome for me is that I sit either alone or sit at a table with no one I know. At a cocktail-style dinner, I can freely move around, and if I get stuck into a conversation with someone whose utterances I cannot understand, then I can quickly (though politely!) cut ties and walk somewhere else.

As luck would have it, Ken Goldberg actually asked his lab members and lab alumni to get together right before the banquet, so that we could walk over (and, presumably, get a table). That was a HUGE relief! I think Ken anticipated this, since he told us a story about how he once got trapped into a table of strangers.

The conference banquet was amazing. Wow. It was a multiple-course dinner, and before each round of food, we had performance artists put on a show. The food included:

• Bread and butter
• Lots of wine
• Smoked salmon
• Guinea fowl (!!)
• Beet and maple yogurt
• Craquelin choux pastry

I ate everything on my plate. The quality was outstanding. The post-dessert coffee was nice because I still had to do some grading for 182/282A.

I knew people at the table I was sitting at, and we got into lively conversations about a variety of topics in robotics and automation. We wrapped up with a discussion of the economic warfare between the United States and China due to Trump’s ill-advised trade wars. I hope this is clear to readers of this post, but I am very, very anti-trade war.

To be clear, perhaps the main reason why I enjoyed the banquet was because I was actually at a table for which I knew some of the people and could participate in the conversation with the help of judiciously-placed sign language interpreters. If those attributes were not present, then my enjoyment level would have been substantially lower.

Note that this was the only main part of ICRA serving essentially an unlimited amount of alcoholic beverages. ICRA certainly did not hold back on making the banquet memorable.

Day 3

For the third main day, Raquel Urtasun gave the plenary. She went through a LOT of material quickly. Normally I am never able to follow talks of this style, but surprisingly, I think I remember more content from Urtasun’s talk than the other two plenaries. This may have been because I read Driverless: Intelligent Cars and the Road Ahead last year, so I knew some of the technical background on self-driving cars.

The second of three poster sessions today was special to me, because that was when I presented our Risk Averse Robust Adversarial Reinforcement Learning (RARARL) paper. The poster session lasted about 1.5 hours, and here is what happened:

• I tried to be active, saying “hi” to the people who came by, and shaking lots of hands. If someone was reading the poster for more than two seconds, I initiated a conversation.
• For those who I talked with, most were just asking for a general, one-minute overview of the project. Fair enough, I think this is standard.
• A handful asked for more specific details on the reinforcement learning algorithm and benchmarks that we were using.
• One person asked me why I did not have a cochlear implant.

After the poster session ended, I felt relieved, thinking that I had done as good a job as I could have. I got the “lunch box”⁵ and ate by the poster pod. Then, the sign language interpreters and I walked over to the next (and final) keynote session.

This one did not go well. The speaker’s accent was so strong that we gave up and left after five minutes. We could have tried the other keynote (remember, at ICRA, there are two keynotes per day, which happen at the same time slot), but I saw from the speaker’s name and country of origin that we would likely run into a similar problem.

Throughout the conference, I was spending lots of time looking at the sponsor exhibits. One of them was The MIT Press, which was selling books. This is my weak spot, and on the third day of the conference, I had to buy a book: The Deep Learning Revolution: Artificial Intelligence Meets Human Intelligence by Terrence J. Sejnowski. I probably know a lot of the Deep Learning history since it’s been in many other books in some form, usually along the lines of: “Hinton, LeCun, and Bengio did Deep Learning when no one else thought it was worth pursuing”. The advantage of Sejnowski’s book is understanding the perspective of a professor who performs research in this area.⁶ Sejnowski, unlike AI reporters, was actually in the subset of academia that developed Deep Learning.

We wrapped up the main part of ICRA with a farewell reception. The reception was in a smaller location than the welcome reception, and the food consisted mostly of cheese and bread dishes. Each of us got one ticket for an alcoholic beverage. The food was good, no doubt about that; I tried to get some whenever the line seemed to subside. The farewell reception was super crowded. I was only able to get a quick conversation with one person there. She was an RIT student. I introduced myself, and told her that I knew people from her school and had visited there several times. It also helped that I remembered reading her poster (about lifelong learning) at a poster session.

This year, I am sure that ICRA spent a LOT more money on the sit-down conference banquet than for the other two receptions. The conference banquet was so outstanding that the receptions, if anything, were almost a letdown. To be clear, I was not upset or disappointed in any way. This is a tradeoff that a conference has to consider. We can’t have everything in tip-top quality, because otherwise registration fees would go through the roof.

After the reception, I ate dinner at a Japanese noodle place. Then it was back to the AirBnB to finalize CS 182/282A grading. I was constantly on Slack and communicating with the other overworked GSIs for the course. We finally finished grading at 2:30AM in Eastern Standard Time.

I slept in, and when I woke up, I realized I was not in shape to go running due to exhaustion from lack of sleep over grading. I ate brunch at a unique chocolate-style cafe with “chocolate BBQ” all over the menu. I then gathered my stuff, went to the Montréal-Pierre Elliott Trudeau International Airport, bought some duty-free Swiss dark chocolate⁷ and … that was it!

Reflections

Upon leaving the conference, I graded myself as “mediocre-to-okay” with respect to networking with other attendees. I still find it very, very difficult to network at conferences, which is one of the main reasons why we have these events. It is hard to initiate conversations with people at receptions and poster sessions. There were a few times that I successfully started conversations, mostly for people who were from Berkeley. Still, for much of the poster session time, I probably spent too much time walking around aimlessly and getting food.

Overall, the food was decent at the poster sessions. I would rate the poster session food higher for Brisbane, since this year, there were too many pastries and dessert-style food. It seemed like at every poster session, the conference attendees were able to entirely finish the fruit-based dish (e.g., strawberries, melons, pineapples, etc.) but left behind much of the sweets and pastries. The banquet, though, was outstanding, and by far and away the most memorable part of the conference, and better than anything Brisbane could offer.

Last year, it seemed like there was some more governmental influence to help recruit people from the conference. We had speakers from the Australian government at ICRA. I suspect that Brisbane, and Australia more generally, has to work harder than Canada does to recruit people and stop a “brain drain.” Canada can probably recruit Americans — or foreigners educated in America — for faculty positions, and has the additional advantage of containing the origin of modern Deep Learning in Toronto.

One interesting note was that there were three people at ICRA (who I did not know beforehand) who introduced themselves to me and said that they knew about my blog. One was an incoming Berkeley PhD student, a second was a recent Berkeley PhD graduate, and a third was from somewhere else, un-affiliated with Berkeley. That was nice!

I know I am privileged in that I don’t have to worry about visa issues. I hope things will be better for students who need visas in the near future.

The sign language interpreters did a great job, and I would be happy to work with them at future conferences.

The iPhone photos I took at ICRA 2019 are hosted at my Flickr account, so that I can remember what the conference was like.

ICRA 2020 is in Paris next year. It will be interesting to see how it differs from the one in Montréal and what tradeoffs the conference organizers decide to pursue.

I will obviously do my best to get a (first-author) paper there. If so, then I promise I will write excessively detailed posts for at least all the main conference days. Stay tuned!