Robots are general purpose machines. Current robotics technology makes them mechanically capable of performing a wide variety of operations. In principle, it is possible to reconfigure a single robot so that it can be used for a range of different tasks; in practice, most commercial users of robots find it uneconomic to reconfigure robots in any situations which cannot be simply demonstrated by ``lead-through'' methods, because of the programming expenses involved. The problem with robots is that although they are general purpose machines, it is very expensive to instruct them for complex tasks.
Commercial robots can perform tasks only when they have been given a precise and detailed set of instructions for carrying out those tasks. The greatest challenge for current robotics research is to implement, in a computer system, the kind of ``intelligence'' that will enable robots to be more easily instructable. In human terms, robots need to be less stupid [Pug82]. The ways in which they are particularly stupid include the inability to act on goal oriented instructions, to plan sequences of actions, to learn from their mistakes, or to understand the world around them; international research in robotics has been addressing these problems since the 1960's.
Research toward providing intelligence for robots is motivated both by the long term commercial benefits of more intelligent robots [Bla86], and by the challenges of robotics for those involved in the wider field of Artificial Intelligence research. Those challenges result from the fact that robotics requires the interface of a computer system to the physical world by way of sensors and actuators. Much early A.I. research dealt either with domains which were purely synthetic (such as computer programming, or chessboards), or with ``toy worlds'', which are simplified computer models of very restricted real world situations.
Whereas robotics imitates the behaviour of intelligent animals by performing the intelligent connection of perception to action, synthetic problems involve no perception or action, and are thereby simplified to a point at which the real problems of robotics disappear. Toy world problems can involve real perception, but act on such a simplified version of the world that many real world problems do not occur. This avoidance of the real world has been criticised by Brady [Bra85a], and by Hayes [Hay83], who suggest that future artificial intelligence research should be linked to problems in robotics.
Brooks [Bro86] has proposed robotics as a starting point for A.I. research, for evolutionary reasons. He points out that higher intellectual functions have only developed in animals over the last few thousand years, once the ability to act in the physical world was well established. He argues that an understanding of the real world as a participant in it is a necessary prerequisite to the kind of ``common sense'' which current A.I. research finds great difficulty in achieving.