James Marshall

In a realistic environment, a robot is continually flooded with multiple streams of perceptual information. In order to use this information effectively for determining actions, the robot must have the ability to focus its attention on the most relevant features of the world by making abstractions and categorizing its experiences. Based on these categories, it must be able to anticipate new experiences and predict how the world will change over time, possibly as a result of its own actions. Most importantly, this process of learning must be driven by the robot's own intrinsic motivation to explore its world, to extend and refine its repertoire of categories, and to choose its own tasks and goals, rather than relying on a set of pre-specified learning tasks. The objective of this project is to design a developmental architecture for a robot that integrates the mechanisms of categorization, prediction, and intrinsic motivation, which will enable the robot to function autonomously in a rich, dynamic environment through an open-ended process of self-driven learning.