While you're here why not look around?
The following material is Copyright © 1997 by Greg Roelofs.
The key point to take home from this session was: test performance at every step of the way, and if it ever drops below about 10 frames per second on your intended minimum target hardware, immediately go back and figure out why. Leaving performance-tuning for the end will almost certainly mean you have to start over from scratch. (A second key point to take home: current authoring tools are almost universally deficient, both in features and as aids to performance-tuning; even the browsers have a long way to go before they achieve minimal conformance with the spec.)
Supposedly the examples (and maybe the slides?) will be available from http://vrml.sgi.com/vrml97/courses/compel/ eventually, but they're not there yet.
In general, though, I found the ``algorithms'' papers to be the most interesting. Their quality was comparable to papers presented at SIGGRAPH, not surprising given that the ACM was a cosponsor. Standouts included John Edwards and Chris Hand's MaPS navigation device, a downward-pointing ``eye in the sky'' camera displaying a real-time, moving, 3D map on a virtual ``tricorder''; and an automated level-of-detail generator called Lodestar, by Dieter Schmalstieg. The latter removes an arbitrary number of vertices (and the resulting null polygons) from a complex model, leaving behind an excellent representation of the original object. (VRML 2 supports, as an efficiency aid, multiple levels of detail, switchable by means of proximity sensors or visibility sensors or whatever; the Lodestar algorithm would allow much easier generation of such worlds.)
The other panels covered multi-user technology, avatars, and user interface design. I found them principally notable for the repeated calls for compatibility--of browsers with the specification, of browser user interfaces with each other (``navigation sucks''), and of various multi-user aspects (including avatars) across platforms and worlds. Note that VRML 2.0 has no support for multiple users; that's being addressed by things like Universal Avatars (semantics or avatar layer), Living Worlds (syntax or VRML interface layer), Open Community (API layer), and VRTP (wire protocol layer).
Aside from the always flashy SGI demos of their Cosmo Worlds authoring tool, the most memorable displays were those by Newfire and Paragraph. The former were showing off their upcoming Heat browser (since renamed Torch) a very fast software VRML renderer based on BSP trees. It was being used to display some Doom worlds and Quake avatars at about 640x480 and around 10 fps. Paragraph was demoing an alpha version of Matchman, a tool for animating avatars; the favored swimsuit model was very lifelike in her movements, although the primary behaviors (walking, running) were canned and almost certainly derive from motion-capture data (Polhemus sensors).
The SGI-sponsored dinner at the Monterey Bay Aquarium, aside from the luxury of the location itself, was notable primarily for the opportunity to eat seafood in the presence of numerous live sea critters--nyuk nyuk. SGI also had several O2s scattered around the building, running an accurate VRML model of the Monterey Bay itself. Tuna, jellyfish, gulls and an ROV were all virtually present in the model.
Right now the VRML phenomenon feels a lot like the very early days of the World Wide Web, except that this time more companies are aware of the potential for a huge explosion in interest, if not profitability. (As we all know, about the only folks making money off the Web are those selling content-creation tools or advertising space at a few popular sites, or those who got an IPO out at the right time...)
The browsers are very strongly reminiscent of the original NCSA Mosaic, in the sense of being functional but not highly optimized. It seems clear that companies like Newfire hope to pull the same coup Netscape did and become the de facto standard among VRML browsers. Let's just hope these companies don't similarly fragment the VRML standard. (For the record, Newfire claims to be implementing their BSP stuff via VRML 2 prototypes, which is the accepted method for extending the language. Other browsers will simply ignore the additional information.) A personal peeve is that almost none of them support the required PNG image format yet; of more than half a dozen browsers I've tested, only the Irix version of Cosmo Player does so far.
Meanwhile the authoring tools are pretty weak; Cosmo Worlds is probably the most complete, but it's only available for SGI's Irix workstations, and even it doesn't do much to help the user optimize his or her worlds automatically. IDS, Dimension X, VREAM and others are joining the fray, however, and no one doubts that there will be considerable improvement over the next few months.