Linear/angular rigid body damping affected by the timestep?

[kermado]

The only thing that I could find in the documentation was here, which suggests that the damping force is calculated by multiplying the velocity by the coefficient. However, looking at bodyCoreComputeUnconstrainedVelocity() in the source code, I can see that damping is applied by taking:

dampedVelocity = undampedVelocity * max(0, (1.0 - damping * dt))

This feels odd, since the magnitude of the damping is now affected by the timestep. A smaller timestep results in less damping. This explains the behavior that I've been seeing. Is there a reason why the damping is multiplied by the timestep? It feels more intuitive to treat the damping as a percentage, so that 0 corresponds to no damping and 1 corresponds to setting the velocity to zero each frame.

[vreutskyy]

I guess it actually was an attempt to make it timestep independent. If you just compute the damping like

dampedVelocity = undampedVelocity * damping

the effect will directly depend on the timestep - halving the timestep will make a rigid body slowdown twice faster.

But of course, just scaling the damping coefficient with the timestep won't do the trick. The correct formula should be something like:

dampedVelocity = undampedVelocity * exp(-damping * dt)

Then the damping coefficient would mean the percent of velocity a rigid body loses per second. With any timestep.

Though, maybe they just didn't want to calculate the exponential function, so used this kind of 'approximation'.

[kermado]

Thanks, this makes sense. I don't think the equation you gave means that the damping coefficient represents the percentage of velocity lost after one second though. I plotted the equation and a coefficient of 1.0 results in a velocity of ~37% of its original after one second.

I'm going to modify the damping to use the equation: dampedVelocity = undampedVelocity * 0.5^(damping * dt). This way, the damping coefficient specifies 1/t, where t is the half-life (the time for the body to lose half its velocity). This should work with any timestep and I think it makes the coefficient a little more intuitive.

I'm approximating the power function using a quartic polynomial, which is faster and provides precision of about 5 decimal places for coefficients ranging from 0 to 1000.

[vreutskyy]

Yes, sorry. That was exponential decay equation (~37% is 1/e)