jaxtransform3d

Idea

This library is still experimental and a lot of things are subject to change, e.g., the name. The core idea is to create a version of dfki-ric/pytransform3d that is JIT-compiled, executable on GPU, differentiable, and inherently vectorized.

Citation

If you use jaxtransform3d for a scientific publication, I would appreciate citation of the following paper for now, since jaxtransform3d is a spin-off of pytransform3d:

Fabisch, A. (2019). pytransform3d: 3D Transformations for Python. Journal of Open Source Software, 4(33), 1159, https://doi.org/10.21105/joss.01159

Maybe not so FAQ

Why should it be inherently vectorized?

The decision is not final yet. As long as it is not too complicated to write the code to accept any-dimensional input arrays I will do it. There are only a few exceptions: rotation matrix normalization methods only accept one rotation matrix. If it gets to complicated or I see significant performance issues in comparison to vmapping the code, I will abandon this idea. The advantage over using vmap is that I don’t have to call vmap multiple times when I want to vectorize multiple times (e.g., a list of a list of a list of rotation matrices).

Which features will this library have?

So far the library contains operations to work with transformation matrices and dual quaternions: normalization, inversion, composition, application to a vector, logarithmic and exponential map. It will stay like this. I will most likely add more operations, e.g., specific Jacobians, sampling, operations with uncertainty, etc. However, unlike pytransform3d, this library will not have any visualization or plotting features. It will not have an equivalent of the TransformManager to manage graphs of transformations. The library will be mostly compatible with pytransform3d though. This library will never get any Euler angle related functions. Nobody wants to differentiate through Euler angle related code.

The library does not have a real user guide. How do I start?

To understand basic concepts, look at pytransform3d’s documentation: https://dfki-ric.github.io/pytransform3d/. This library is documented through the API documentation and the examples. The examples should give you an idea of how to use it with JAX’ vectorization, differentiation, and JIT-compilation functionality.