[ODE] repeatability in collisions?

[Victor B. Zordan]

Hello.

I see a similar question about deterministic behavior,
mine is slightly related:

I am running an optimizer to find control inputs for a
behavior, which requires multiple runs of my ODE sim...

I am wondering if anyone else has noticed problems with
the ODE collision code having slight discrepancies from
run to run.  That is, starting from the same initial
state (pos,rot,lin vel,ang vel for all bodies) and running
for a while with collisions turned on yields a slightly
different state across runs.  However, with collisions
turned off, the end state from run to run is identical.

I am trying to determine whether a repeatability bug is
known and if there is a workaround (for example, reset
some random seed or ???)

Any input welcome!  Thanks,
Victor