Because we couldn't release the code and data we've also prepared a small example implementation which I hope will prove useful to those trying to implement the method:

github.com/gouruiyu/Con...

Briefly, Control Operators provide a way of encoding arbitrary inputs to Neural Networks. This is particularly useful for game development as you can't be sure what the structure of input data will be like. For full explanation check out this blog post we prepared:

theorangeduck.com/page/impleme...

Yeah - very much possible if you drop down to C or use something like numba/cython. The Python overhead is just insanely large - often 100x slower than the C equivalent for something like quaternion maths, even in places where you can use numpy.

Made a bare-bones Motion Matching implementation to test it: github.com/orangeduck/G...

Sadly I had to drop to 30fps because Python is so painfully slow even something as simple as Forward Kinematics can take several milliseconds. Eeek.

Does your font have a name? And would you mind if I included a version of it in my article as an appendix 🙂?

Oh awesome didn't know about these!

Really interesting... scary how much the inner robot is capable of without a conscious control loop. I can simple envision the words I want to appear on the screen and it will type them out for me without any visual feedback or semantic understanding of what a "key", "keyboard" or "screen" is.

Big engines are using the FBX SDK (it's crappy and slow but the only sane way to get data out of an FBX file).

For prototypes I usually end up writing a maya script to dump out the raw mesh data and bind pose into a binary format I can handle. Usually something like this: github.com/orangeduck/G...