How might one design a PopBeast which can survive in such worlds? The answer depends on how much information one is willing to program it with. The PopBeast I describe below already ``knows'' the basic properties of keys, doors and other objects when it starts life. However, a truly artificial intelligence ought to be able to learn these for itself, just by wandering around the world, playing with things, and noting regularities. At first sight, it might seem easy to build something which does this - which notices that when you grab a rock it doesn't move, but when you grab a key it does, and therefore keys can be picked up. Indeed, in Eden, building such a program would not be too difficult. Objects always behave in the same way; there are very few objects, and very few actions one can perform; and perceptual data (vision, anyway) is always complete.
In the real world, objects have many more properties and interact in many more ways. Building an AI program that can decide what properties and interactions to take note of is not trivial - look up Bongard problems in Hofstadter's Gödel, Escher, Bach (Penguin 1981: PSY KH:H 67) if you are not convinced.
Moreover, real world objects differ in all sorts of inessential ways (consider the difference between two poodles); there are many more things you can do with them; and no two situations are ever alike. We are a long way from solving the problem of machine learning, and I have circumvented it by pre-programming PopBeast with knowledge of its world's basic physics. This is a big cop-out, since the ability to learn should probably be central to every intelligence, with everything else depending from it.