Motion generation from world descriptions: the level of required redundancy

An overview is given of the experimental analysis of motion planning from comparison of two snapshots of the world: initial scene and goal scene. It is commonly believed that from comparison of these two images the researcher can deduce the plan of actions and, eventually, the program of motion (e.g. in robotics). Indeed, in many particular cases the existence of a knowledge inverse operator can be proven. If so, the process of plan generation can be done automatically. The feasibility of solving the problem of automatic plan generation is considered. These questions are addressed by considering several man-chair situations. An effort is made to analyze this process experimentally and to trace the whole set of required computational procedures, as well as the required structure of knowledge representation. It is concluded that difference generation requires a substantial amount of knowledge which is not represented explicitly within the interpreted image description. Supervised learning is one of the possible ways for filling the lists of primitive rules and metarules. However, using redundant thesaural descriptions seems to be more promising