Learning and Planning with Probabilistic Relational Rules

Abstract

The Problem: Our research involves learning models of world action dynamics, which can then be used to construct plans to reach a wide range of goals. The work is applied to simulated worlds, such as the blocks-world environment shown in Figure 1. Figure 1: A three-dimensional blocks-world simulation built with the OpenDynamics toolkit [7]. The world consists of a table, blocks of roughly uniform size and mass, and a robotic hand that is moved by simulated motors. Motivation: Robust robotic control in complex worlds is a challenging problem. Hand-engineering a solution is difficult and time-consuming. Developing techniques that will allow robots to gather knowledge about the world and use it to design their own control strategies seems like a reasonable alternative. Previous Work: We represent world action dynamics using probabilistic planning rules. Figure refrelrules-fig shows two rules that model actions that can be performed by the robotic arm in the blocks world of Figure 1. Such rules enable us to take advantage of the inherent structure found in many uncertain, complex environments by making the following assumptions about the world: • Frame Assumption: When an agent takes an action in a world, anything not explicitly changed by that action stays the same. • Object Abstraction Assumption: The world is made up of objects, and the effects of actions on