Using genjax to invert graphics and dynamics.

[kamiradi]

Hi Genjax community!

I am new to the PPL paradigm and eager to learn by building a project in this area. For my robotics project, I plan to use the inverse graphics machinery for pose estimation as described in the original Bayes3D paper. While I intended to run the code provided in the paper, I would like to reimplement the ideas so I can integrate dynamics into the framework. For this purpose, I will be using Drake, which features its own rendering based on OpenGL and supports differentiable dynamics.

I am seeking pointers or recommendations for available tutorials and relevant sections of the Bayes3D codebase, which I found quite challenging to navigate. My goal is to develop this machinery alongside the Sequential Monte Carlo inference. I intend to document the process and contribute any useful infrastructure back to the Genjax community. Additionally, the project will leverage the differentiable capabilities of JAX and Drake, making it an exciting research endeavor.

Would this be of interest to you? I’ll wait for initial responses before I outline the specifics.

Thank you!