Mesh collision not working properly

[cherylwang20]

Hi,

Thanks for the great package! I am a student and I was trying to use obj2mjcf to convert a beaker.obj model into sub-meshes to use as a hollow object in mujoco.

I managed it separate it into 28 convex hulls, however, when I load the mesh into the env I have with a table and robotic arm, it penetrates the table and has its COM on the side.

I have attached the initial setup:

and after free falling from a height:

As a reference, the box and the mug model from MuJoCo are behaving fine.

Please see here for the xml code:
https://github.com/cherylwang20/mj_envs/blob/3e73f238eea48cffed0423d7fcd604b467838b4a/robohive/envs/arms/ur10e/scene_gripper.xml

Any help is greatly appreciated! Thank you.

[kevinzakka]

Can you share a ZIP file containing the original mesh, the decomposition and the scene.xml with all the assets? I can take a closer look :)

[cherylwang20]

Can you share a ZIP file containing the original mesh, the decomposition, and the scene.xml with all the assets? I can take a closer look :)

Hi Kevin,

Thanks for the quick reply. Here is the zip file.
beaker_file.zip

If you drag the scene_gripper.xml into mujoco, it should reproduce what I have. The beaker.mtl and beaker.obj are the original mesh files. the decomposition is in the folder 'beaker'.

Thank you again and let me now if you need anything else.

[kevinzakka]

Taking a look, thanks @cherylwang20!

[kevinzakka]

Hi @cherylwang20, the XMLs you sent were super disorganized so I stripped everything down to just the beaker and the table and tweaked the physics to get better behavior.

I found that 15 convex hulls are sufficient to approximate the beaker well. Here's the obj2mjcf command I ran:

obj2mjcf --obj-dir . --obj-filter beaker --save-mjcf --compile-model --decompose --overwrite --coacd-args.max-convex-hull 15

In terms of changes I made to the XML:

1. Switched to elliptic cones.
2. Changed collision condim to 6.
3. Stiffen contacts by reducing solref[0] (note it should not go below 2 x your timestep).
4. Toggled multiccd to get more stable contacts between the table surface and the beaker.

Cheers!

beaker_scene.zip

[cherylwang20]

@kevinzakka Thanks for helping me look into this problem. And apologies for the inconvenience in XML file organization.

Thank you again for the prompt support and the development of this package!