This is a very old 1990 paper, and gives a flavor as to what the quality of research was like back in those days. This paper also mentions "attention" and must have inspired some recent work on attention models. (Though this paper hasn't been well-cited in recent years.)
Their modeling assumptions vary slightly from the standard MDP stuff I'm used to (again, welcome to 1990 research) but it can be converted over, and more importantly, they also maximize the sum of discounted rewards as I would. Yay.
This work is about trying to do this while also having perceptual aliasing. Unfortunately, it's not clear to me what they mean by that. My guess is that by that term, they mean multiple real-world states get mapped to a common internal representation (i.e. "featurization") and thus they get aliased and have the same utility (value) function, even though the actual states might vary considerably in real-world value.
To overcome this, they propose an algorithm which lets them perform several perceptual actions to adjust the state configuration, doing so in a way that changes the sensors, but not the actual world state. This results in several different viewpoints of the world. Then they perform a "real" action. This makes sense. (Unfortunately, they don't describe how to choose the most informative set of perceptual actions w.r.t. a cost budget.)
What doesn't totally make sense is the need to have a "lion" state as they describe it. The lion, roughly speaking, is an internal state that best matches the outside, real world state. I think this is like what happens in POMDPs? Maybe they're equivalent but the literature at that time doesn't explain it clearly enough.
This seems more about how to choose the proper internal representation of the state and not necessarily the "best" way to actually achieve that.