To better reproduce the experiments of the article 3D Steerable CNNs: Learning Rotationally Equivariant Features in Volumetric Data go to this commit
For point cloud I recommand to use this code: https://github.com/e3nn/e3nn
For voxel I don't know which one is better