Goal-oriented Flexible Action Generation of Robot by Neural Network with Contextual Consideration from Past to Future

Abstract

There have been progresses of understanding the mechanism of human perception and action based on computa- tional theory such as predictive coding. Especially by experiments using a robot with a recurrent neural network (RNN), the process of adaptation to an environment and action generation according to a goal is explained based on predictive coding. However, goal-oriented flexible action generation within multiple action options has not been well investigated. This research aims to show how goal-oriented flexible action of robot is generated based on pre- diction error minimization. By implementing an RNN with adaptive mechanism considering prediction error from past to future to the robot, action search considering both the environment and goal was confirmed. In addition, untrained action which efficiently achieves the goal was observed. This result shows that goal-oriented flexible action generation can be explained by considering prediction error of both past environment and future goal.