An industrial robot routinely carrying out an assembly or welding task is an impressive sight. More important, when operated within its design conditions it is a reliable production machine which - depending on the manufacturing process being automated - is relatively quick to bring into operation and can often repay its capital cost within a year or two. Yet first impressions can be deceptive: if the workpieces deviate somewhat in size or position, or, worse; if a gripper slips or a feeder jams the whole system may halt and look very unimpressive indeed. This is mainly because the sum total of the system’s knowledge is simply a list of a few variables describing a sequence of positions in space; the means of moving from one to the next; how to react to a few input signals; and how to give a few output commands to associated machines. The acquisition, orderly retention and effective use of knowledge are the crucial missing techniques whose inclusion over the coming years will transform today’s industrial robot into a truly robotic system embodying the ‘intelligent connection of perception to action’. The use of computers to implement these techniques is the domain of Artificial Intelligence (AI) (machine intelligence). Evidently, it is an essential ingredient in the future development of robotics; yet the relationship between AI practitioners and robotics engineers has been an uneasy one ever since the two disciplines were born.