[ODE] LCP solution methods

[Jon Watte]

> I'm sorry that I'm not familiar enough with ODE to make some tests 
> and have to ask for explanations instead of looking for them in the 
> code. But discussions may sparkle some ideas, which is mutually 
> beneficial. Besides, I'm essentially interested in a better method 
> than ODE has (that is, if there's no tricks I don't know about that 
> ODE uses to make SOR converge faster)
 
I believe the method used by commercial toolkits (including the very expensive ones) is similar to the ODE StepFast solver with over-relaxation. If someone on this list was math-savvy enough to know of an even better method, I'm pretty sure they'd either have put it into ODE already, or else be working hard trying to sell a very expensive commercial library, and not reply...

Personally, I understand the math, but I'm not comfortable enough with it to formulate my own solutions, so I probably can't help you more than that. From my experience on embedded solutions, though, you have to realize that constraints really do exist, and design around those constraints, rather than trying to brute-force something "smarter". Paraphrased: If you can't do 5,000 objects on the GameBoy DS, then design your game for no more than 2,000, or come up with clever ways of only solving the small subset of objects that the player can currently see or get close to, and leaving the others in limbo (not moving). If you're doing non-game solvers for embedded devices, where the design can't change, well, you've got your work cut out for you :-)

Cheers,

            / h+