I did a few experiments regarding that in 5.8 build from source.
- First of all, seems like with
p.Resim.AllowRewindToResimulatedFrames = false
p.Chaos.BubbleResim.Enable true
np2.Quantization.Enabled true
no physics LOD, no pawns being physics LOD focus
I can run the whole scene with 80-100 connected objects on potato client PC without any performance issues. So I decided to try working only with resim objects because 80-100 objects is not gonna be in real scenario anyway.
- I’m testing this scene. Vehicle Pawn in resim + 4 wheels in resim connected to Vehicle Pawn via constraints in Vehicle Pawn BP. On top of that 3 the same cargo physics objects with resim
- Test #1 with full resim. If ping exceeds 250 ms, panels can go through the vehicle, stuck in between bed and wheels, generate some force that leads to local tornado. Usually it fixes itself after some time, but it’s getting worse with larger ping, and in general still noticeable even on lower pings, even when client drivers the vehicle on a flat surface with constant input.
I tried:
a) np2.Resim.RuntimeCorrectionEnabled = true (with and without velocity correction, with and without CorrectConnectedBodies) - it’s much worse and could lead to unfixable desync
b) increase these values p.Chaos.BubbleResim.ExtraThicknessForCorrectedParticle, p.Chaos.BubbleResim.ExtraSweptDistanceForCorrectedParticle, looks a little bit better, but nothing major. (I also tried without BubbleResim, the same thing)
-
Test #2 with full predictive interpolation instead. Here things look interesting, CorrectConnectedBodies set to true makes the system quite robust, but with a cost of significant delay in vehicle control. However as soon as some Resim actor (character for example) jumps on the vehicle, it goes crazy and fails.
-
Test #3 with mix - vehicle in resim, panels in predictive interpolation. Panels are moving behind the cargo bay, but no local tornado.
Test #3 example:
So I wonder are there any planning stuff with resim with connected bodies? Or I miss some important options to test?
PS: In general seems like that 5.7 was much more stable in that specific scenario than 5.8, both with Test#1 and Test #3