Action Feasibility Learning with Cell-Based Multi-Object Representation for Task and Motion Planning

Abstract

This paper proposes a description method and a deep neural network classifier for action feasibility learning for robotic tasks. The proposed method is able to deal with multiple three-dimensional obstacles and continuous action parameters. These advantages significantly enhance the effectiveness of the method in an unplanned environment. Thus, we expect our method can significantly increase the usability of task and motion planning in daily-life applications.