Why isn’t decreasing the integration time step an effective method of increasing stability in bullet? Maybe it is, but it did not seem to be in my experience.
Interpolating the normals doesn't improve simulation stability exactly, it reduces the effect of driving over edges.
There's no way we can afford the CPU time of a simulation step of 1000Hz. We were at 120Hz, but we're shipping at 60Hz, since we're not getting collision problems at that rate. Your timestep should probably be set based on the fastest speed of your objects and the size of those objects to eliminate tunneling.
To increase the stiffness of the constraint solver, increasing btContactSolverInfo.m_numIterations will have more effect than the timestep. You might also try a different constraint solver (Featherstone, or analytic.) As I understand it though, fighting stiffness like this is a losing battle.
Note, I haven't tried any of this: the default 10 constraint iterations, and the sequential impulse constraint solver were enough for my needs. My simulation only involves one rigid body per car, so I don't have a stiff simulation.
I'm interested at ANY code, any hint about that
some time ago I tried to model some Mc Pherson and Double Wishbones based chassis using Bullet constraints and "my own" suspensions. It was highly unstable, the whole thing deintegrated on bumps
I was wondering if soft bodies could help, or where the (stable but unrealistic) raytracing based VehicleDemo could be modified, but I ran out of motivation and, possibly, talent. I'll look at that, thanks
It sounds like your simulation is too stiff. I don't think soft bodies will help, because they still need to be solved by the constraint solver AFAIK. I think you can still get realistic simulation from a raycast vehicle though.