I'm torn between: Learning systems (POMDP, RL, etc) for planning, Quantum computing basics, or jumping whole hog into the a learning-from-scratch type course tr

I'm torn between: Learning systems (POMDP, RL, etc) for planning, Quantum computing basics, or jumping whole hog into the a learning-from-scratch type course track. With RL+Planning, my goal is to eventually be able to tradeoff classical estimation and planning with end-to-end systems, and use the best parts of either for a better-than-either pipeline. At the jobs I've had I've seen a weird false dichotomy that produces decent results once committed, but there's always this tension that "the other way" is better, producing weird politics. Most recently, a classical planning pipeline was completely broken by an ML-based estimation of orientation of obstacles. On inspection they were using quite possibly the worst estimator you could use, and ignoring all kinds of good data, because (I think) they were counting on perfect "measurements" from the ML-based vision system. That kind of thing must happen all the time - get stuck in a local minimum of "tune just a little better" when a good filter on top of a noisy estimator could solve all your problems. There's probably a million such examples in planning systems too.