A hypothetical restoreManifoldsFromPoints for partial state persistence

[micimize]

Hi – I'm working on a project where partial physics state serialization and restoration is desirable, hopefully with deterministic results. We're using Godot with bullet3 physics, forked with a few modifications, and are successfully saving all state beyond that which is cached by the solver & collision cache, which I am now investigating.

At this point I've seen most of the threads out there about the challenges posed (i.e. @erwincoumans 's forum comment from 2011, #2982 (overlapping pairs determinism setting)).

So, after much digging around, this is my current thinking:

1. It seems like the only physics state cached outside of rigid body data between steps is the list of contacts for each manifold.
2. Restoring the manifolds and/or full cache state directly might be tedious for us, because we don't have the same body pointers between save & reload.
3. However, we could more easily add an API like getManifoldPoints()/restoreManifoldsFromPoints() in our fork, which would:
   a. Persist a mapping of body pairs to a list of manifold points
   b. Restore them by calling dispatcher->getNewManifold for each entry, and re-add the points with addManifoldPoint

Does this make sense, or is there other persistent state that might effect determinism that I'm not accounting for here, or some other potential issue?

Thanks for your time,
Michael

[micimize]

@erwincoumans sorry to bother you, but I might give this idea a try soon for use in the avalon benchmark/training environment and was hoping to get some expert feedback on whether it is feasible before burning a few days on it.