The water-to-land transition around 350 million years ago had several notable consequences. One is a 10-fold increase in relative brain size over ancestral water-dwelling animals. Another, which our own group has shown, is that the volume of space where something can be visually detected increases by a factor of one million, due to the higher transparency of air to light compared to water and typically larger eyes. Because of the vast amount of plant life on land as well as landscape barriers to visual sightlines, animals on land observe distant and near objects, which modulate between “in-view” and “hidden” as an animal moves. If the target is sentient, predicting the target’s behavior while the target is hidden becomes challenging. One way this uncertainty can be reduced is by having an internal model of the target. Inasmuch as the target cares about the perceiver—for example, if the perceiver is a common prey for the target—that model will necessarily contain a reference to the oneself (most generically, for a predator, this model would include that if it sees the prey it will try to get closer and eat it). Thus, prediction of the hunter by the hunted requires that the hunted have a model of itself. Both aspects of the internal model---how it predicts its target, and in what way the self is contained in the model of the other, is needed for effective planning. We describe results from a set of experiments with robots interacting with mammals in a predator-prey dynamic through a partially occluded space where we have made some initial progress on how to tackle this set of issues.