The paper describes the prototype HAP - heuristic assembly planner and tries to explain current ideas about how to extend the present search algorithms using principles of AI - machine learning. HAP planner uses AO* based heuristic search through AND/OR graph representing all possible assembly sequences. The main problem of this approach was shown to be its computational complexity, which was somehow limited by developing efficient algorithm for generating feasible assembly sequences. Still, when the number of such sequences are to big, the search may not be efficient. To cope with this, we present the idea about how the system can learn the "assembly rules" from previous planning and how such a knowledge can be used when searching for the optimal assembly sequence of a new assembly.