I used to write physics engines for animation, back in the 1990s when nobody had one that worked right. That required reading books on nonlinear differential equations and getting consulting from experts at Stanford. I had to learn about quaternions. I had more of a classical computer science education - number theory, mathematical logic, combinatorics, proof of correctness - but not enough number crunching.

Before that, I'd worked on automatic theorem proving and proof of correctness. I still like Boyer-Moore theory. I recently revived the old 1970s-1992 Boyer-Moore theorem prover and put a working version on Github. It's fun to run that again; it's a thousand times faster than it was in the early 1980s.

If you do anything serious with graphics, you need to understand 4x4 matrix transformations throughly. I have the whole shelf of Graphics Gems books, and they're mostly math. At one point I rewrote many of the C code in C++, and got rid of their start-at-one arrays. (The original was Graphics Gems in FORTRAN, and the C version used a horrible hack to make arrays start at 1.)

I didn't know enough filter theory when we were doing the DARPA Grand Challenge. We had a lot of trouble integrating the GPS and AHRS data into a good position and orientation. We had about 3 degrees of heading noise, which kept messing up the map-making function. We really need 3D SLAM, but didn't know how.

Now I need more math to understand machine learning.

I'm also looking at designing a specialized switching power supply for the antique Teletypes I restore. You can get enough energy from a USB port to drive the big selector magnet if you use and store it properly. Fortunately I can get LTSpice to do most of the number crunching.

I think I've used all the math I was ever taught. And I'm not really into math.