Object Feature Extraction and Tool Using for Robots with Recurrent Neural Network Model

Construction of an efficient perception and action mechanism is the most crucial issue for creation of intelligent robots. Our objective is to apply the robot’s own experience for developing the robot’s perception and action mechanism based on human development. In this talk, I present my work on development of robot’s perception mechanism based on affordance theory and motion generation using tools. Concerning perception, affordance theory suggests that humans’ cognition is based on perception of invariant features of the environment [1] [2]. Two types of invariant features exist: structural invariants and transformational invariants. Structural invariants are features that remain constant through motor interaction and transformational invariants are features that describe the change of the environment. The first half of the presentation is on using the robot’s pushing motion to tabletop objects to extract features that represent structural invariants and transformational invariants. The second half of the presentation is on motion generation with tools. Tool use is one of the largest distinctions of primates from other beings. The ability to use tools as if they are part of the body is called tool-body assimilation [3]. Tool-body assimilation experiments with a humanoid robot generating goal-oriented motions with and without a tool are presented.