Highlights of Robotics: Science and Systems 2012

For me, the most interesting work at the conference were two related papers, one from Ed Olson and another from Niko Sünderhauf.

A Generic Approach for Robust Probabilistic Estimation with Graphical Models. Niko Sünderhauf, Peter Protzel.

They both address the question of how to deal with multi-modal observation errors, such as data association uncertainty in SLAM, wheel slip in odometry, multi-path effects in GPS, etc. Ed Olson’s paper uses a very simple trick, namely replacing Gaussian uncertainty with a max-mixture of Gaussians, letting the estimator handle multi-modal observation pdfs. The nice thing about this is that it’s incredibly simple to do. With minor changes, you can use this with existing estimators like iSAM. As shown in the figure above, it’s very effective, easily handling a high percentage of incorrect loop closing constraints when vanilla estimators fail completely. The caveat is that there are clearly local minima in this new approach, and Ed mentioned at his poster that the natural thing might be to revisit SGD-like optimization techniques. I think  the current results rely on the metric estimates being not-too-bad at loop closure. However, lots of potential.

The paper by Sünderhauf and Protzel tackles basically the same problem, but uses explicit switching variables to handle the multi-modality. I missed this talk, and haven’t yet had a chance to read it in detail, but it seems to also achieve excellent results, albeit perhaps it’s more expensive to run and more complex to implement.

Overall I’m very excited about this work, it’s a simple idea which seems to solve one of the gremlins that prevented SLAM from being truly reliable. Appearance-based loop closure techniques developed in the last few years could deliver loop closures with 99% precision (e.g. my own work on FAB-MAP and many related systems), but eventually you still get false positives and the system was in trouble. Robust estimation was sorely needed. It’s kind of amazing no one has done this before. I vaguley remember some related work on human pose tracking, but for SLAM it seems to be new. Particle filter devotees may ask what’s new here – couldn’t particle filters do all this years ago? Well, I think these new techniques have hope of working long-term, where particle filters fail due to diversity loss.

Anyway, nice to see some people coming back and cleaning up the not-really-solved problems with SLAM. The community overdosed on SLAM and moved away for a few years, which is healthy, but the problem was not truly solved and it’s nice to see people coming back to do the extra work needed to make it really work robustly.

Breifly, some more papers which stood out. Some of these I don’t know enough about to fully judge, but they all seemed interesting:

Robust Object Grasping using Force Compliant Motion Primitives Moslem Kazemi, Jean-Sebastien Valois, J. Andrew Bagnell, Nancy Pollard

Walking and running on yielding and fluidizing ground Feifei Qian et al.

Visual Route Recognition with a Handful of Bits Michael Milford

On Stochastic Optimal Control and Reinforcement Learning by Approximate Inference Konrad Rawlik, Marc Toussaint, Sethu Vijayakumar

Affine trajectory deformation for redundant manipulators Quang-Cuong Pham, Yoshihiko Nakamura

Outside of the main conference, there was lots of good discussion in the quadrotor workshop. The invited talks as ever were also a real highlight. Andrew Howard stole the show describing  the SpaceX Dragon mission docking with the International Space Station using his vision system. Finally Charlie Kemp gave a great early-career talk entitled “Mobile Manipulation for Healthcare”. There’s no paper, but here’s a video showing one of the projects.

If you’re only interested in robotics in general and not the minute details of conferences, stop reading now. I want to finish with a few meta-comments on how the conference itself was run. In short, I think the changes to the RSS format this year were pretty much all improvements.

The biggest change was that the workshops were no longer on separate days, but instead ran before lunch each day of the conference. This allowed the main body of the conference to be four days instead of three. I really liked this new arrangement, it simultaneously increased participation in the workshops and gave the main program more breathing room. This had the knock-on effect of allowing the 5-minute talks to be more spread out, which was a huge improvement. Last year there were some sessions with twelve 5-minute talks in a row. My head asplode. This year 5-minute talks were in groups of at most five, which worked much better. The total number of papers presented actually went up substantially this year  (from 40 to 60), but it felt less pressurized. To accommodate the shift, most talks moved to the 5-minute-plus-electronic-poster format, with only award candidates having full length presentations. There were some complaints that this format is too light-weight, but personally I think the combination works. The electronic posters really are dramatically better than paper ones, so a teaser talk plus an in-depth discussion at the poster session doesn’t feel like such a compromise any more. For a single-track conference it makes sense. Though with integrated workshops, RSS is now almost multi-track by the back door…

Finally, I think the program committee did a good job on the reviewing. RSS is defined by its very high reviewing bar and by a broad focus including contributions from outside the robotics mainstream. While this sounds great, in the last couple of years I felt like the conference was losing its way a bit, partly by becoming too broad, and partly because the high reviewing bar was producing a bias towards a certain kind of math-heavy paper that often seemed to me more technical flourish than real advance. Some of the most interesting content was being pushed into workshops, outside the main program. I was very glad to see what I felt was a much more balanced program this year.