If I remember correctly, Apple's Voxel Net (~2018 neural network, used for doing machine learning on sparse lidar point-cloud data) uses a simple "spatial hash" for this purpose, using the coordinate of the voxel as a key.
Yeah, hashing for spatial data is a fairly common technique (and not invented for that paper). It totally makes sense for applications where there is a need to index sparse data in a potentially very large or even unbounded physical space. There's some hashes that are specialized for keys that are 3-dimensional points or voxel coordinates, but almost any off-the-shelf hash will do if applied to say, the concatenated indices of a voxel in a given coordinate frame.