The Dynamics section contained a number of papers on dealing with rigid and soft body dynamics. Because most of the papers deals with the math involved in simulating physics for these mechanisms, I won't provide a lot of detail, but there were a couple of talks that were particularly interesting.
This paper, delivered by Pixar (and the only one that I saw delivered clearly on Apple's Keynote product), discussed how to make cloth correctly untangle itself after having been crumpled, folded, or spindled.
The problem is that during simulations, cloth is often simulated as a thing layer very close to the skin, however, since most simulations have invisible flaws in them, it is not uncommon to have cloth (or even skin) interpenetrate itself, thus creating a tangled mess. To resolve this, animation programmers have used tricks mostly involving the automatic insertion of repulsion forces when a piece of cloth gets too close to itself, however this only works until the cloth actually intersects itself. If, for some reason, the cloth interpenetrates before the simulation notices it, these additional forces will act to keep the cloth tangled.
In this paper, the folks from Pixar suggest the use of global intersection analysis to be able to determine (without requiring history) when a cloth is interpenetrating and to right it.
I'm sure the paper is more interesting than the speaker, who was pretty dry, but the results were amazingly good.
Nonconvex Rigid Bodies with Stacking
The students from Stanford that presented this paper had some interesting demonstrations to show involving very large numbers of non-convex rigid bodies interacting with collision and stacking.
It wasn't clear to me if this was truly a real-time system, but it appears that by changing the order of operations in determining collisions and contact made them able to run the simulation with fewer calculations and get a much faster result. If you have a chance to look at their "1,000 rings falling on a bed of spikes" demo, it's rather nice.