For rendering, I wanted to go one step (but not more than one step) beyond the flat-shaded polygons that are typical for physics demos. I wanted shadows, but I have never been a big fan of sharp edges, such as with shadow volumes or even shadow maps. Soft shadows is really what enables a whole new level of realism.
I worked on an ambient occlusion project for Ageia a couple of years ago, just before they got acquired by Nvidia and I think that's a really good alternative to real global illumination. My approach at the time was to compute ambient oclcusion in 3D on the second generation PhysX hardware (that was never released) and dynamically update low-res light maps for the parts of the scene that changed. It had good potential, but shortly after Crysis was released and the screen-space methods (SSAO) started taking off. I've always been curious about SSAO ever since but never got around to implement one, so I thought this was a good opportunity.
I found this article, which is a really good introduction. I didn't end up using exactly that method, but it's quite similar. Vertex position and normals are rendered to an FBO, then the occlusion is computed to another FBO and the final image is rendering to the framebuffer using deferred lighting. During the final pass occlusion values are blurred, using only samples in the same plane. Ideally one would want to do this with a gauss kernel, but I only do it horizontally and vertically to save some shader cycles. I'm still on my three year old MacBook Pro with an ATI X1600... Additionally I'm also blurring the normals a tiny bit during the final pass to soften the edges.
There's not an awful lot of time spent on the rendering part, but I'm quite happy with it for the time being. Here's a video demonstrating the scene with and without ambient occlusion.