Question about benchmarks about maniskill HOT 4 CLOSED

PhysX internals are a bit out of my scope of understanding, @fbxiang might understand it better but tbh PhysX is not super well documented.

https://nvidia-omniverse.github.io/PhysX/physx/5.3.1/docs/Simulation.html is the docs about the whole substepping thing, could not find more details on dt.

from maniskill.

PhysX is known to do less solver iterations with smaller dt. And smaller time steps in general makes a lot of simulation problems easier to solve.

from maniskill.

Thank you for the kind comment!

So in ManiSkill, there are two temporal levels of simulation. There is the underlying higher frequency physics simulation powered by physx, and then there is the lower frequency control simulation where each step of control is one applied action (target positions/velocities are applied to active joints).

In ManiSkill, we take one control step and apply the given actions and then take sim_freq / control_freq physics steps.

High sim frequency affects the dt configuration, saying how much time passes when taking a physics step. If sim freq is higher, then less time is taken between physics steps and I imagine this means less solving is needed.

As a result you can observe the following if you add --save-video to your two commands (and lower number of envs to something storeable in a video)

Top video is sim, control freq of 500, 100, bottom video is 100, 20 (ratios are kept the same)

Each frame of the video is a single control step, comprised of 100 / 20 = 500 / 100 = constant 5 physics steps. Hence with higher sim freq it looks smoother/slower per control step since less time is passing, lower sim freq more time passes between control steps (and solver is less accurate)

from maniskill.

Awesome, thanks for the insight!

So it sounds like as the physics step time (dt) is increased, PhysX takes a longer amount of wall time to do a single physics step (potentially due to the constraint solver iterations being dynamic and dependent on dt?). That's quite interesting, any further insight on the PhysX side would be appreciated.

Closing the issue for now, thanks again!

from maniskill.