`mjContact.vert` bugs for `flexcomp`-`geom` collisions

[kurtenkera]

Intro

Hi!

I am an Australian PhD researcher at the University of Queensland, and I use MuJoCo for my robotics research.

My setup

MuJoCo version: 3.2.3 API: Python Architecture: x86_64 OS: Ubuntu 22.04.1 LTS

What's happening? What did you expect?

I'm trying to use MuJoCo's elasticity plugin to accurately simulate contact forces on soft, deformable robots. However, I've noticed that mjContact.vert returns incorrect flex vertex ids when simulating a simple flexcomp soft cube colliding with a plane geom.

Note that the cube that I have considered in my tests was meshed in gmsh. The picture below shows the vertex ids of the meshed cube in GMSH. In particular, note that the base of the cube has vertex ids 1, 2, 3, 4 and 9. From personal experiments, I've noticed that MuJoCo uses the same vertex id labelling scheme as GMSH, however all ids are based from 0 in MuJoCo (i.e., each id in GMSH is subtracted by 1 for zero-indexing).

When I set the value of euler of my flexcomp cube to non-trivial values, bugs arise.

Steps for reproduction

1. Download the contents of softbox.zip.
2. In the same directory create a new empty Python file called minimal.py and copy and paste the Python code below into your script. Make sure you adjust the xml_filepath variable in line 4 accordingly!
3. Set euler in line 25 of softbox.xml to the following values: "euler = 1 0 0", "euler = 2 0 0", "euler = 45 0 0", "euler = 0 1 0". For each case, run minimal.py. In each case, the box is dropped from a height and ultimately lands on flex vertex ids 0, 1, 2, 3 and 8 (or 1, 2, 3, 4 and 9 are the GMSH ids as mentioned earlier). However, if you inspect the terminal output, there are some flex vertex ids that are repeated, and we get the incorrect flex vertex ids involved in the contact between the flexcomp and plane in each case.
4. Confusingly, if you uncomment the if condition in line 14 of minimal.py and repeat step 3 for all cases, we get the correct output/flex vertex ids for all cases. What is causing this!?
5. Repeat step 3 for the following values: "euler = 90 0 0", "euler = 90 90 0", "euler = 90 0 90" and "euler = 90 0 180". You will get the correct flex vertex ids regardless of whether the if statement in line 14 of minimal.py is commented out or not. So it seems 90-degree rotations are handled correctly, but anything else produces erroneous bugs?

Minimal model for reproduction

The following zip folder contains the mesh of my soft cube - softbox.msh - and a minimal MJCF file called softbox.xml: softbox.zip.

Please run the Python code below following the instructions above to reproduce the bugs.

Code required for reproduction

import mujoco
import os

xml_filepath = "path/to/softbox.xml"
model = mujoco.MjModel.from_xml_path(xml_filepath)  
data = mujoco.MjData(model)
duration = 3.0 
n_steps = int(duration/model.opt.timestep) #timestep = 0.0005

for i in range(n_steps):

    mujoco.mj_step(model, data)

    # if i == n_steps - 1:
    print("t = ", data.time, "\n")
    for con_id, con in enumerate(data.contact):
        print("Contact #", con_id+1)
        print("con.vert = ", con.vert, "\n")

Confirmations

• [X] I searched the latest documentation thoroughly before posting.
• [X] I searched previous Issues and Discussions, I am certain this has not been raised before.