One of the best labs I ever did in university (this was circa 1970), was to create an Initial Program Load (bootstrap) program for a virtual IBM System/360, which copied punch cards, using no OS assistance. (Modern zSeries I/O is so complicated that you'd never ask people to write it bare, but the 360 was just barely manageable for that.)

I think that wrangling with a computer in the absence of an OS is the only way for a person to understand what the OS is actually doing for you. It's all very nice to study synchronization schemes or memory management, but the methodologies all seem pretty abstract until you have to face a real computer.