ok, if you manage to pull objects apart of your 3d depth eye. (which cant be that hard alot of the time)
any identified object, can have a backpropagation of its relative values of its sorrounding environment, to its movement.
that way you can now 'pretend' your other things.
then you can randomize a calculus transform of relative variables to the environment again to the motor outputs of the creature... and if you can pretend what the other things do, you can simulate as fast as you want and develop a real physical solution to the problem its having with other moving objects.
you can supply more and more power, to the randomizer, i dont know how good random can develop successful behaviour... depends on how you score it definitely.