Given a (potentially nasty, nonconvex) mesh, automatically creates an SDF file that describes that object. The visual geometry is pointed to the input mesh, and the collision geometry is pointed to a convex decomposition of the input mesh (as performed by trimesh and vhacd.

This has been designed to interoperate well with Drake, but doesn't strictly need it.

For just basic conversion:

argparse lxml numpy trimesh

and you need testVHACD on your PATH. I use the procedure from this script to get it.

The visualization utility requires Drake.

To perform the conversion (here, on an example mesh with a 1000x downscaling):

python generate_sdf_from_mesh.py test_data/bowl_6p25in.obj --scale 0.001 --preview

Use --help to see arguments; you can tune some VHACD parameters to influence the quality and complexity of the resulting decomposition. The defaults are tuned to create lightweight approximations of relative simple concave geometry; if you have more complex geometry, try increasing resolution, max # of hulls, and verts per hull. In particular, these settings have been practically successful for me:

python generate_sdf_from_mesh.py <path_to_obj> --maxNumVerticesPerCH 32 --maxhulls 64  --resolution 1000000 --preview

If you have Drake installed, you can inspect the resulting SDF (use the menu at the right to show/hide the visual and collision mesh components):

python inspect_sdf_in_meshcat.py test_data/bowl_6p25in.sdf

Both can be invoked with --help to see additional arguments.