Development of a 3D Convolution Neural Network for Classifying Robot Assembly form Features

The voxel-based 3D convolution neural network (3D CNN) proposed in this paper classifies form features to decide candidate assembly directions for automated robot assembly planning. It can classify not only form features of a part but also its candidate assembly directions that are needed for the following assembly planning procedures. In the implemented automated robot assembly planning system, it will take the place of the current rule-based form feature classification module. The 3D CNN uses classification classes that integrate both form features and their directions to assist candidate assembly directions to the following component ordering and robot assembly planning procedures in the automated robot assembly system. This study generated 3D CAD models for each form feature class and converted them into voxel models for the training of the 3D CNN. This study also contrasted rule-based classification methods with voxel-based CNN and evaluated the advantages and disadvantages of each method.