Random number generators are kind of interesting, because they can't exist
They do exist but they're difficult because you can't generate them with logic. You need to harvest a bunch of enviromental information (time of day, core temperature, # of key strokes in the last week etc) and combine them to produce some form of seed. The problem is that any can be non-random under certain circumstances and therefore all of them potentially could be non-random. The answer to that is to use alot of them.

If you want to generate the exact same sequence of "random" numbers for a test, you can supply a specific seed and get the exact same sequence over and over. I've never seen a need for this, but I recognize that it could be useful.
We use it alot for regression testing AI algorithms. They're inherently random so having a truly random sequence makes them effectively untestable. Having a predictable pseudo random sequence solves that problem. I've never seen a need for it in production code though.