louxinxiang
commented
11 hours ago I have the same question. I used my own dexterous hand model, which is watertight but not in the shape of capsules/boxes. I found it quiet strange that even hand model didn't contact with object meshes, the penetration term was not zero but a big number. Probably because of this reason, many candidates after 6k iterations didn't contact with the object meshes. Have you met the same problem before?Looking forward to your reply!
@mzhmxzh @wrc042
I tried to visualize how TorchSDF generates SDF for a mesh, but I found it performs bad. When I specify a mesh (eg. ffdistal) and calculate the distances of voxel point cloud of a 0.20.20.2 cube, it turns out there are many 'leaky cones' shot out from this mesh, which means they are wrongly considered as inner side points. I guess it's because the mesh isn't watertight. After I fixed my mesh with manifoldplus in your repository, the mesh become watertight, but the problem get nothing better. Then I relooked your mjcf file, and find that you regard very link of your collison model as boxes and capsules, then it is simple and easy be calculated by TorchSDF! So I'd like to ask if there are any tools to convert mesh to boxes and capsules or you carefully designed it manually? Plus, as you only use boxes and capsules, then your torchSDF function is only for boxes? cause you heuristicly calculated SDF for capsules. Is it true?