Yes, you could, but I don't think anyone would call doing ellipsoids without rotation realistic. You can get away with it for spheres if they have no friction, and it's easier to ignore as a sphere looks the same when rotated, but it would just look odd for ellipsoids. And even then there are multiple problems. Your collision response cannot take place in this scaled space, so you need to scale everything back if you want a bounce and not just a hit-test. You have to scale the world multiple times, once for each ellipsoid with different geometry or [fixed] rotation. And if there are any spheres or circles in your world they become ellipsoids, so you have the same ellipsoid-sphere problem you were scaling to avoid.PaulCunningham wrote:There's no 'magnitudes harder' if you don't need to consider rotations. You simply scale your world geom so your ellipse becomes a unit sphere.
The problem is while polygon-polygon intersection problems (and so hit-tests) can be solved with linear equations, and polygon-sphere or sphere-sphere need quadratic equations, ellipsoid-sphere, ellipsoid-ellipsoid, ellipsoid sphere (and annoyingly cylinder-edge) problems result in quartic equations. They are in my book magnitudes harder, as although solvable algebraically the solution is far more difficult and subject to far worse problems with rounding/fp errors. When I had to include cylinders in a physics engine they had realistic physics when rolling on a flat surface but when colliding with another object used a prismatic approximation.
I prefer subdivision in cases like this. First if you have a complex or detailed world then you need to do sweep tests against each surface you might collide with, and decide which one you hit first. Second doing a sweep test for the end collision is harder: the points where it hits lie on a circle, so you have to do a ray-circle test and solve a quadratic. None of this is too difficult but it's a lot of extra logic and code when all you need to do is change the time by 1/2, 1/4, 1/8, ... and re-run your hit-test code to determine when the collision happened as precisely as you need. For a single object the difference in performance is not an issue...based on the same principal and uses sweep tests so you don't have to use sub division if you don't want to.






