Learning to predict action outcomes in continuous, relational environments

We present a method for predicting action outcomes in unstructured environments with variable numbers of participants and hidden relationships between them. For example, when pouring flour from a cup into a mixing bowl, important relations must exist between the cup and the bowl. The action Pour(x, y) might depend on the precondition Above(x, y). How well the predicate Above actually predicts action success often depends on complicated world dynamics and perhaps other objects in the scene. While such predicates are commonly hand-crafted, we present in this paper a method for learning physically grounded predicates directly from the continuous data. In this manner, an agent´s own developmental experience can drive its world representations. Here, we learn such representations as ensembles (or forests) of probability trees using the Spatiotemporal Multidimensional Relational Framework (SMRF). By reasoning about individual objects, SMRF trees allow us to focus attention on action parameters while still considering other objects in a scene. We demonstrate our method on three simulated problems. Two are in a blocks world with gravity: one predicting the tipping direction of a balance scale, inspired by Siegler´s classic cognitive psychology work, and the second predicting the success of dropping one block on another. The third problem predicts the success of passing a ball in a soccer domain. In these tasks, we show an ability to scale prediction to scenes with more objects than are present in the training data.