1998-12-15
11:15 at HG G26.1

On the significance of materials for intelligent robots

Rolf Pfeifer

AI Lab, Computer Science Department, U Zurich/CH

Early approaches to understanding intelligence have abstracted from physical properties of individual organisms. The generally accepted assumption was that intelligence can be studied at the level of algorithms which is why for many years t ial properties, and the relation to control (for example, most robots almost exclusively employ rigid materials). The goal of this talk is to investigate - or rather to raise - some of the issues involved. After a short introduction into the basic framework of embodiment, a number of examples are discussed: - the passive dynamic walker and the problem of "exploiting dynamics vs. control" - natural and artificial muscles and the problem of "degrees of freedom" - the relation between complexity of neural control and material properties: the "face robot" vs. the "Bit" approach ("Bit" is the name of a robot doll). - "Computational properties" of materials: an extension of the theory of computation - "Morpho-functional machines", i.e. machines that change their shape depending on the situation and perform different functions (e.g. amoeba robots) - Sensors of various types: the case of skin sensors and the problem of concept acquisition These issues will be elaborated using a number of case studies. For example, the human arm is an enormously complex system with many degrees of freedom - a control engineer's nightmare. But muscles and tendons have elastic properties that impose cons In conclusion, if we are to understand brain function - and intelligence - we must understand the physical setup, the morphology, and the materials. The big challenge then is to bring together task-environment (what the robot is supposed to do and it The talk will conclude with a very brief overview of the research projects in our laboratory to provide additional context for the issues raised.