Investigation of a single object shape for efficient learning in bin picking of multiple types of objects

Abstract

In recent years, many studies have used Convolutional Neural Networks (CNN) as an approach to automate bin picking tasks by robots. In these previous studies, all types of objects in a bin were used as training data to optimize CNN parameters. Therefore, as the number of types of objects in a bin increases under the condition that the total number of training data is retained, CNN is not sufficiently optimized. In this study, we propose a learning method of CNN to achieve a bin picking task for multi-types of objects. Unlike previous learning method using multi-types of objects, we use a single type of object with a well-designed shape to obtain training data. It is true that when an object is used for training, then the learning for the corresponding object proceeds very well. However, the training does not contribute so much to the leaning of other types of objects. we expect the CNN to learn many types of grasping methods simultaneously by the training data of the single well-designed object. In that case, the total number of training data for each type of objects in a bin can be retained even if the number of types of objects increases. To verify the idea, we construct 12 different CNN models which are trained by different types of objects. Through simulations and robot experiments, bin picking tasks to pick multi-types of objects were performed using those CNN models. As a result, the training method which uses a complex-shaped object achieved higher grasping success rate than the training method which uses a primitive-shaped object. Moreover, the training method which uses a complex-shaped object achieved higher grasping success rate than the previous training method which uses all types of objects in a bin.