Development of humanoid robot Mybot-KSR and Its behavior selection scheme for human-robot interaction

Abstract

This paper presents mechanical and electrical design of humanoid robot Mybot-KSR for HRI research. Mechanical design of integrated joint module with torque control ability is proposed. The integrated joint module is utilized to develop the arm of the humanoid robot. The humanoid robot is also equipped with main control PC for stand-alone execution of posture control and HRI algorithm. Four kinds of subcontroller boards are developed for joint motor control, hand control, foot pressure sensor control, and communication control, respectively. They are developed with MCU abstraction layer concept which modularizes MCU firmware and hardware and increases reusability of the design. Joint controller board is custom-designed for torque controllable integrated joint module. This paper also proposes a novel layered confabulation architecture for hierarchical behavior planning. The confabulation based planning architecture yields a sequence of behaviors which achieves the goal of given task while considering past experiences in the memory. To achieve this, confabulation process, based on the memory replicating human thought process was merged into a goal-directed tree searching algorithm. Confabulation based behavior planner is composed of semantic memory, confabulation memory and process, and STRIPS style tree search based planning algorithm. An ontology based semantic memory using situated affordance concept is proposed. Situated objects and behaviors in the semantic memory are quantified and each transformed into assumed fact symbols and conclusion symbol of confabulation module and used in the confabulation based reasoning process. In STRIPS style planner, a set of situated objects is expressed with nodes of search tree as states representing environmental status. The tree searching algorithm of the planner draws appropriate behavior order using cogency information of behaviors obtained from confabulation processes as well as behavior costs and heuristic estimation values. This method is applied to the task intelligence for two arms operation of the humanoid robot. In order to plan the behaviors of two arms hierarchically, three confabulation based behavior planning layers are constructed. Abstract behavior planning layer, manipulation behavior planning layer, and primitive behavior planning layer are constructed so that the appropriate sequence of behavior to achieve the goals can be drawn. With demonstrations of two scenarios, beverage serving, and cereal and milk serving with a real humanoid robot, the proposed architecture showed its effectiveness.