View from convention center balcony

Size

Anchorage welcomes its new robot overlords

There were roughly 1,600 registered attendees, so the event was spread out over two big convention centers. All of Anchorage’s hotels were packed with roboticists. It wasn’t long until local cafetarias were playing into this.

More than 2,000 papers were submitted submitted, of which 857 were accepted. While that’s an acceptance rate of only about 40%, it’s still one heck of a lot of papers! Instead of actual paper proceedings we received a booklet called the “Conference Digest”, containing a single slide for each paper. With six slides on each side of each page, this thing is still about the same size as the full proceedings to SoCG or EuroCG. The actual proceedings came on a DVD, and my tablet PC lacks a DVD drive, so I have not been able to look at them yet. But I presume that there is still a lot of space left on the DVD, so I wonder why there is a 6-page limit on the submitted papers. For the initial versions it makes sense, as this reduces the reviewing load, but why not allow larger final versions?

Diversity

Image from the slides of my talk

To accomodate all these papers there were thirteen parallel tracks; but really, it felt more like thirteen parallel conferences. Robot-building disciplines like mechanical engineering, electrical engineering, and computer science come together with application areas for robotics such as the life sciences, geology, logistics, and manufacturing. It seems like you could present research on just about any topic (“X”) at ICRA, provided you word it the right way (“X using robots”). Heck, I attended one talk that went into anthropology! While it was nice to get a glimpse of what is happening in other fields, and how they all relate to one another, most speakers assumed the audience already had intimate knowledge of their specific field and its acronyms. I felt a bit left out, most of the time, and wondered whether that described half the audience or just me.

Workshops

Willow Garage's PR2 checking us out

On Monday I attended the workshop on Best Practice in 3D Perception and Modeling for Mobile Manipulation. My main take-away was that we’ve come farther than I thought, and that the state-of-the-art is all freely available as open-source libraries. You just download the ROS robot operating system and OpenCV computer vision library, and you’re all set. In a couple of lines of code you can combine a stereo image into a single image where you know the distance of each pixel to the camera. A bunch more lines and you can build up a fully textured 3D model of the world from a stream of these images, captured as the robot moves around.

Of course, that’s when the tricky part starts of determining what all the objects you are seeing are, and what you should do with them. This is where the open problems are. The rest of the workshop had some talks on these. All were interesting techniques for different kinds of special cases, but we are still nowhere near dealing with the full variety of objects humans deal with on a daily basis.

Technical sessions

Tuesday till Thursday were exhausting. A 12-minute talk every 15 minutes from 8:30 to 19:30, except for breaks and the occasional plenary or keynote speakers who had longer talks. The sessions ended earlier on Wednesday because of the conference banquet, but that’s still more than 70 talks I attended. For many of them I only vaguely understood the problem that was being solved, and usually had no idea what the state-of-the-art was and how the presented method improved on it. Still, there were a couple of talks I found really enjoyable.

• My own talk was on pushing a disk with a disk-shaped robot, and it turns out there were two more talks on the subject of pushing (though all in different sessions, for some reason). In A Dipole Field for Object Delivery by Pushing on a Flat Surface, Takeo Igarashi presented a very simple and elegant algorithm for the same problem I looked at, and even showed how it worked with actual Roombas. The algorithm cannot (yet?) deal with obstacles, but I had a nice chat with him afterwards about future work. In Dynamic Pushing Strategies for Dynamically Stable Mobile Manipulators, Pushkar Kolhe (how awesomely appropriate is that name?) studied how a differently designed robot should push or pull in order to exert the most force. It turns out that pulling is actually better than pushing.

• In Adaptive Multi-Robot Coordination: A Game-Theoretic Perspective, Gal Kaminka talked about a new protocol for moving robots to avoid colliding with each other. They could easily have had the paper accepted after just showing through experiment that their method works, and I would have thought nothing special of the talk. Instead, they went the extra mile in using game theory to also show why the method works.

• In Avoiding Zeno’s Paradox in Impulse-Based Rigid Body Simulation, Evan Drumwright explained how physics simulation libraries such as ODE and Bullet get resting contacts wrong, and presents an alternative method that has been implemented in a new physics library called Moby.

• My favorite talk actually won the best paper award in its category. In Gesture-Based Human-Robot Jazz Improvisation, Guy Hoffman presented his jam sessions with Shimon, a marimba-playing robot. Rather than play a preprogrammed piece of music, it improvises based on what the human is playing. The result sounds great, and the addition of a head that head-bangs to the beat was a nice touch. It really looks like it gets into the groove!

Tutorials

On Friday morning I went to the workshop on Guaranteeing Safe Navigation in Dynamic Environments. All great talks on how to avoid collisions among moving obstacles, each illustrated with pretty videos. Jur van den Berg showed an especially impressive video on simulating human movement in crowds.

Friday afternoon I attended the tutorial on Real-Time Planning in Dynamic and Partially-Known Domains. It emphasised how almost any planning problem can be reformulated as finding a path in a graph where edges are labeled with a cost and/or probability. The obvious way to do that is using A* search, but I had no idea how many different variants of A* have been developed to deal with different kinds of problems. Almost a dozen of them were explained, having funky names like Fringe-Saving A*, Lifelong Planning A*, and Anytime Repairing A*.

Homecoming

The A* tutorial at the end was actually the first time during this conference that I saw someone use theorems and complexity theory. What’s perhaps a little disturbing is that that actually made me feel relieved, as if I had come home from a long ardous journey through the wastelands. That’s not to say that the other talks at the conference were bad, but they were rather different from the kinds of talks I’m used to. I think computational geometry may have spoiled me a bit in that regard. I’m used to talks with clear, precise definitions and theoretical analyses. It’s easy to forget that in the “real world” one deals with vague concepts and must rely on experimental validation.

A lot of interesting topics were presented in talks ranging in quality from moderate to excellent. I have to give special mention to Amit Chattopadhyay. When the speakers for the talk 2-Factor Approximation Algorithm for Computing Maximum Independent Set of a Unit Disk Graph” could not attend the conference, he offered to give the talk instead of letting it get cancelled. Not being one of the authors, nor being affiliated with them, he just read their article and their slides a day in advance and gave a talk on-the-fly. Kudos!

Fast-forward sessions

As decided by vote at the business meeting of EuroCG 2009, this year’s EuroCG was the first to have fast-forward sessions. In these short sessions, all conference attendees are together in one room. The speakers of the next two parallel sessions get 60 seconds each (with slides submitted in advance) to promote their talks. There is a coffee or lunch break before the actual parallel sessions start, and then the process repeats. This way, attendees are supposed to be able to make a more informed choice about which talks they want to attend.

This notion stolen from inspired by SIGGRAPH was still a bit unfamiliar to both organizers and speakers. During the first few fast-forward sessions the speakers were asked to go in “column order”, that is, first all the speakers of session A, then all speakers of session B. At the suggestion of several attendees, this was changed to “row order” so that “competing” speakers went one after another. This order was then kept for the rest of the conference.

In both orders, the speaker of session A always came before the parallel speaker of session B. This seemed unfair to me, but no one present seems to have used this to their benefit. More generally, I think most speakers (myself included) did not think the fast-forward through from a strategic point of view. My parallel speaker even had a slide with the title of our paper on it, saying something to the effect of “this guy is speaking in parallel to me, and his talk will surely also be nice”. Certainly a nice gesture , but I’m not entirely sure why someone would do this except to get a smaller audience.

All in all, I quite like the concept. Skimming the proceedings in advance says something of the topics, but not of the speakers. The addition of fast-forward sessions thus gives one this extra variable by which to decide which talks to attend. At the business meeting, the vote was n-3 against 3 in favor of keeping the fast-forward sessions, so they will be used again at EuroCG 2011 held in Morschach, Switzerland.

Keynote speaker

Although all the invited speakers gave excellent talks, the one by Timothy Chan was the most memorable to me. With his inimitable enthusiasm he gave a one-hour talk on Instance-Optimal Geometric Algorithms.

He explained the concept using the example of computing 2D convex hulls. For some, “easy” sets S of n points there are algorithms to compute its h point convex hull in O(n) time, while for other, “hard” point sets different algorithms with a running time of O(n log h) are known to be optimal.

Timothy defined a function H (somewhat similar to entropy) so that H(S) is a number that measures how “hard” it is to compute the convex hull for point set S. More precisely, he proved that there is a Ω(H(S)) lower bound for computing the 2D convex hull in the multilinear decision-tree model, by a simple and elegant adversary argument. He also showed that a slight modification of the Kirkpatrick-Seidel convex-hull algorithm yields a running time of O(H(S)). Thus its running time (up to constant factors) is as good or better than any other 2D convex hull algorithm that does not depend on the order of the points.

Timothy ended the talk with applications of the same technique to other problems, which in some cases did require new algorithms, such as for 3D convex hull. All in all a very impressive talk on some very impressive research, the paper for which can be found on Timothy’s website.