The trouble with tree based structures is they’re expensive to build. Unlike hashing, where for each frame, you can just iterate through every object inserting it into a hash table and check for collisions at the same time.

Also I have a little opengl/C demo that throws ~1k balls around the screen. I could send you the source to if you want more explanation?

A property of AABBs I like is that sibling bounding-boxes can overlap slightly, with little impact on overall utility. Also, when objects move “a little”, AABB extents can be re-adjusted without shuffling objects between nodes, whereas (as you point out) spatial subdivision methods typically requires shuffling moving objects around whenever they cross a boundary

“I use circles for collision detection, and otherwise whenever I can for organization, because the results look more natural than using boxes. If you organize using boxes, your simulation will tend to “align” to the axes, which nature surely does not do.”

I would distinguish between geometry that /participates/ in the simulation (leaf object hulls), and geometry that /optimizes/ the simulation (axis-aligned bounding boxes [AABB]). I’d be curious to understand if you believe use of AABBs could “axis-align” the simulation dynamics, if the AABBs themselves never result in collision responses?

Cheers,
Sam