mujoco Inconsistent contacts information between mjx and mujoco when the geoms are a capsule and a box.

Hi, I've found an inconsistency with the data.contact list between mujoco and mjx. I'm wondering if I've stumbled upon a bug or I've did something wrong.

Here's a minimal example that reproduces the problem:

import mujoco
from mujoco import mjx

XML=r"""
<mujoco>
<option>
  <flag gravity="disable"/>
</option>

<worldbody>
  <body pos="0 0 0">
    <freejoint/>
    <geom type="box" size=".1 .1 .1" name="box" rgba="1 1 1 0.25"/>
    <site pos="0 0.02500001 0.85" size="0.025" rgba="1 0 0 1"/>
    <site pos="0 0.02500001 1.15" size="0.025" rgba="1 0 0 1"/>
  </body>

  <body pos="0 0 1">
    <freejoint/>
    <geom type="capsule" size=".15 .15" name="capsule" rgba="1 1 1 0.25"/>
  </body>
</worldbody>

<contact>
  <pair geom1="box" geom2="capsule" margin="100"/>
</contact>
</mujoco>
"""

mj_model = mujoco.MjModel.from_xml_string(XML)
model = mjx.device_put(mj_model)
data = mjx.make_data(model)
data = mjx.forward(model, data)

print("mjx:")
print(f"contact.dist: {data.contact.dist}")
print(f"contact.pos: {data.contact.pos}")

mj_data = mujoco.MjData(mj_model)
mujoco.mj_forward(mj_model, mj_data)

print("mujoco:")
print(f"contact.dist: {mj_data.contact.dist}")
print(f"contact.pos: {mj_data.contact.pos}")

And here's the output of the above code:

mjx:
contact.dist: [1. 1.]
contact.pos: [[0.         0.02500001 0.85      ]
 [0.         0.02500001 1.15      ]]
mujoco:
contact.dist: [0.6 0.9]
contact.pos: [[0.   0.   0.4 ]
 [0.   0.   0.55]]

The outputs I get show different contact distances and positions between mjx and mujoco. Here are some screenshots showing the contact positions in red dots, left is mjx and right is mujoco:

Also, why does it return two contacts with different positions? It seems to me that the actual midpoint is [0, 0, 0.4], why does it return another midpoint with a different position?

Context:

Operating system: Ubuntu 22.04
Mujoco version: mujoco-mjx-3.1.2

Mar 05 '24 03:03 Manchewable

I can answer the second question. Your geoms are so large with the margin, that the box-capsule collider is colliding the capsule as if it is two spheres (imagine a capsule embedded in a box, these collisions are what you want), therefore the two points returned by MuJoCo are the collision points for the two spheres. @btaba to comment about MJX collisions.

Mar 05 '24 07:03 yuvaltassa

Thanks @yuvaltassa! I've just stumbled upon https://github.com/google-deepmind/mujoco/issues/1325 which I think answers my first question.

To follow up on my second question, why is it that if I only move the capsule along the x or y-axis, it'll return only one contact position as opposed to two?

If I assume that it'll return two different contacts most of the time, and I want to know the contact distance between those two geoms during simulation, would it be right for me to just take the contact with the smaller distance?

Mar 06 '24 02:03 Manchewable

The images above are for MJX or MuJoCo? I'm not sure I know what you're asking without being able to reproduce.

If I assume that it'll return two different contacts most of the time, and I want to know the contact distance between those two geoms during simulation, would it be right for me to just take the contact with the smaller distance?

Yes

Mar 13 '24 09:03 yuvaltassa

Hi @Manchewable , thanks for the bug report. It turns out margin wasn't really implemented for capsule-box, and so we just pushed out a change to raise an error if margin is present for those collisions in MJX. It isn't expected that the capsule-box impl in MJX will give exactly the same results as in MJ as explained in https://github.com/google-deepmind/mujoco/issues/1325

Mar 14 '24 23:03 btaba

mujoco mujoco copied to clipboard

Inconsistent contacts information between mjx and mujoco when the geoms are a capsule and a box.

mujoco
mujoco copied to clipboard