Deep Imitation Learning for Broom-Manipulation Tasks Using Small-Sized Training Data

It is important for robots to learn the usage of tools and support humans in aging societies. It is expected for robots possible to imitate human skills of tool manipulation properly using a deep neural network, although a huge amount of training data may be required. In this paper, a target human-like task of cleaning dust using several types of brooms with a robot arm is considered. A learning system that can reduce the amount of training data is proposed. The novelty of the proposed system is the ability to estimate the initial parameters of a deep neural network based on the shape of the broom and data stored from previous experience. Furthermore, the system changes the number of learning layers in the deep neural network depending on the broom shape. Results of experiences show the effectiveness in reducing the amount of training data.