Minimal reproduction of very unstable contact forces for robotiq gripper simulation

[RussTedrake]

What happened?

I've carved out a minimal example of the poor performance we've been seeing in the contact modeling with the robotiq gripper. I thought it was specific to SAP, but i've included two models here: one with SAP + coupler constraints, and another with TAMSI + bushing elements. Both exhibit massive chattering in contact force directions. The problem persists even if the timestep is reduced considerably.

The reproduction is in this notebook. (I can't upload the ipynb directly)

You can view the behavior in this meshcat recording.

Version

master

What operating system are you using?

Ubuntu 22.04

What installation option are you using?

compiled from source code using Bazel

Relevant log output

No response

[RussTedrake]

I've ported a version of the first test to C++ for you here: https://github.com/RussTedrake/drake/pull/new/robotiq_minimal_repro

It should be trivial to switch in the second model (the last cell of the notebook) if you want to look more closely at that one, too. I still recommend running this notebook first.

[xuchenhan-tri]

The link to the c++ repro seems to be wrong. Here is a correction:

https://github.com/RussTedrake/drake/tree/robotiq_minimal_repro

[xuchenhan-tri]

It seems to me that this is more of a model issue than a solver issue. The collision geometry of the finger tip is messed up. One of the finger tip is modeled as a matrix of tiny spheres for point collision and the other finger tip is modeled as a box, all using point contact. The jittering is likely caused by the well-known point contact issue between two box faces.

After manually curating the asset and replacing all the tiny spheres with hydroelastic, the contact force is a lot more stable: https://github.com/xuchenhan-tri/drake/tree/robotiq

[RussTedrake]

I'm sorry that this was still open. Alejandro and I iterated on it and resolved it similarly (and did even better with hydro)