Designing grounded agents: From robocup to the real-world

This paper discusses the nature and role of "grounding" in designing programs for controlling autonomous mobile robots. Since its inception, artificial intelligence has been plagued by problems of scaling and brittleness. A fundamental problem impeding the development of artificial intelligence is our dependence on grounding agents by design. That is, currently agents tend to be grounded by their designer's understanding of the world, task, and robot. However, little (if any) of the knowledge of "how to ground" is embedded in the artificial agent. Consequently, brittle, purpose-built systems result. This paper explores how the intellectual burden of grounding can be shifted from the programmer to the program by designing robots capable of grounding themselves. An overview of a grounding oriented design methodology (Go-Design) is presented - an initial step towards the longer-term objective of developing autonomous grounding capabilities.