The problem of peg-in-hole assembly remains to be an important subject of continued investigation mainly due to its wide applicability into virtually all practical assembly tasks, Consequently, some effective methods are developed and, in particular, the insertion of a peg into a chamfered hole can readily be carried out with the aid of a passive remote-center-compliance (RCC) wrist installed in an industrial robot. Nevertheless, for the case of nonchamfered and high-precision part mating only a few applications using special hardware such as a vibratory equipment are successfully implemented. This study propose a practical solution based on the mapping conception under the following conditions; (1) chamferless (2) small clearance and (3) minimum use of the special hardware. Researches based on the mapping relation between the signal from the force sensor and the corresponding action can be divided broadly into two classes; (a) the mapping relation is predetermined through the task analysis (b) the mapping relation is obtained by learning through iterative executions of the task. The former approach is unsuitable for practical usage because it depends mainly on the task model and has its limitation due to the difference between the actual environment and the task model. The letter, however, can be applied even to the unanalyzable portion of the task because the mapping strategy can be acquired directly from the actual environment without task analysis. Also, the latter can easily be applied to the various task environments with the different parts.
This research propose a practical and efficient method for acquiring the mapping relation between sensing data and corrective motions of a robot. A binary database is used to represent the mapping relation where experiences are memorized in pairs of a sensor signal and a corresponding action. The proposed method is based on the fact that the input-output relations acquired during the learning stage are memorized in a self-organizing way. In addition, it has the ability to autonomously establish the correct mapping as well as to accelerate the learning by use of the priori knowledge obtained from the partial analysis of the task. The proposed method will formulate the algorithm for acquiring the mapping relation, the database for representing the input-output relation, and the application method for all stages of assembly. To evaluate performance of the proposed method, simulation studies were made over a wide range of assembly conditions and various uncertainties. The results show that the proposed algorithm is robust to various uncertainties and is still effective in a wide range of assembly conditions.
Finally, in order to verify the effectiveness of the proposed method, a series of experiments taking the actual environment into consideration were executed in all stages of assembly task. The experimental results show that the sensory signal-to-robot action mapping can be acquired effectively and, consequently, the assembly task can be performed successfully.