Put-In-Box task generated from multiple discrete tasks by humanoid robot using deep learning

Abstract

We propose a robot manipulation model using deep neural network (DNN) which is able to perform multiple shorter sequential tasks in series to complete longer sequential task. Execution of multiple tasks is a necessity for robot's usage variety, but recent researches of robot manipulation with DNN focus on single tasks and do not focus generation of multiple tasks in series. Our model extracts image features using autoencoder and recurrent neural network model is used to generate “Put-In-Box ”task from three divided phases: open the box, pick up the object and put it into the box, and close the box. The three phases are trained separately, but the model successfully switches motions using extracted image features to perform “Put-In-Box ”.