World Models: A Definition and Development Roadmap.

"World models," defined as internal simulators that learn an environment's structure and dynamics, have become a central but often inconsistently understood concept in AI. Researchers across various subfields, from reinforcement learning to robotics, are developing systems they call world models without a unified definition or clear development path. This article aims to provide clarity by offering a scientific definition of world models. It delves into their crucial technical aspects, discussing what these models should predict and how they ought to be constructed. Furthermore, the paper presents a staged roadmap for the effective development of world models. This roadmap seeks to establish a common understanding and guide future research efforts, fostering consensus in this rapidly evolving and highly debated area of artificial intelligence.