[throwaway37585]

This behavior wouldn’t necessarily transfer to the real world because the real world has costs (e.g. energy utilization and hardware damage, both very important in nature) which are not always accurately reflected in these simulations. It brings to mind the example where an agent learned how to make a cheetah “run” while repeatedly banging its head on the ground, which wouldn’t work in the real world for obvious reasons.

[adrianmonk]

Also, for a human there is cognitive load in moving around. If you are safe where you are now and nothing significant changes, it's mentally easier to stay still instead of re-evaluating everything constantly. And this frees your brain to better plan your next move, so it's advantageous. For an AI, CPU power isn't as scarce.

And even with a computer playing a video game (not the real world), your joystick hand gets tired.