The world modelers project: objectives and simulator architecture

Machine learning has long sought to construct complete, autonomous learning systems that start with general inference rules and learning techniques, and gradually acquire complex skills and knowledge through continuous interaction with an information-rich external environment [59]. The World Modelers project provides a simplified artificial environment — a continuous three-dimensional physical model of the world — to facilitate the design, implementation and testing of integrated learning systems. Thus, the world simulator can be viewed as a tool that enables researchers to investigate significant intermediate steps between the single-function learning systems of today (systems largely incapable of multi-task learning, attention focusing, experiment formulation, etc.) and the real-world autonomous learning robot of tomorrow.