Machine Intelligence & Robotic Control, Vol. 1, No. 2, 79–86 (1999) Paper Behavior Learning of Human-Friendly Robots by Symbolic Teaching* Naoyuki Kubota † , Shuzo Yamaji † , Fumio Kojima ‡ , and Toshio Fukuda § Abstract: This paper deals with behavior learning of human-friendly robots by human symbolic teaching. The mobile robot has an internal model for its behavior criteria and acquires human teaching model based on the behav- ior criteria. Outputs of human teaching model are used for learning reactive motions such as collision avoidance behavior. The feature of this method is to obtain suitable behaviors through the interaction with environment and symbolic teaching by human intuition. Experimental results show that the robot can acquire collision avoidance behaviors through the interaction with human symbolic teaching in a given environment. Keywords: Human-Friendly Robot, Fuzzy Controller, Behavior Coordinate, Collision Avoidance 1. Introduction H UMAN-FRIENDLY robots are required in various fields including service industry and welfare. Re- cently, various methodologies for robotic control have been discussed in subsumption architecture, behavior-based robotics, and evolutionary robotics [1]–[7]. These con- cepts are based on reactions that living creatures present. Robot’s reactions can be described by production rules, neural networks, and fuzzy inference rules, which are ac- quired by learning in environments. However, a human- friendly robot should acquire its behaviors through inter- action with human in a given environment. Furthermore, the evaluations of human concerning the robot’s behav- iors are different among human operators. This means that a human-friendly robot should acquire behaviors suit- able to a certain human operator. In this study, we dis- cuss a learning method for human-friendly robots. The learning methods can be classified into three types: super- vised learning, unsupervised learning, and reinforcement learning only with the response of success or failure [8]– [12]. If a human operator can give exact teaching data, the robot can acquire a behavior suitable to the human opera- tor. However, it is difficult for the human operator to rep- resent exact numerical teaching data. Actually, symbolic communication such as “turn right,” “go up,” and “stop” is often used in teaching among human operators. In such a case, the robot must understand the meanings of sym- bolic teaching data. Therefore, the robot must build the human teaching model by itself. Here we use symbolic communication to share information between the robot and human operator. Therefore, the robot requires a mapping method from human symbolic information into numerical information for learning various behaviors based on human * Received December 13, 1999; accepted January 20, 2000. † Dept. of Mechanical Engineering, Osaka Institute of Technol- ogy 5-16-1 Omiya, Asahi-ku, Osaka 535-8585, Japan. E-mail: kubota@med.oit.ac.jp ‡ Department of System Function Science, Graduate School of Science and Technology, Kobe University, 1-1 Rokkodai-cho, Nada-ku, Kobe 657-8501, Japan. E-mail: kojima@cs.kobe-u.ac.jp § Center for Cooperative Research in Advanced, Science and Technol- ogy, Dept. of Mechano-Informatics and Systems & Dept. of Mi- cro System Engineering, Nagoya University, 1 Furo-cho, Chikusa-ku, Nagoya 464-8603, Japan. E-mail: fukuda@mein.nagoya-u.ac.jp teaching. Consequently, we propose a behavior learning method based on the mapped information. As one of ba- sic experiments, we focus on a collision-avoiding behavior of the human-friendly mobile robot in this paper. Gener- ally, the robot should take into account various objectives simultaneously, such as collision avoiding and target trac- ing. Therefore, we propose a motion coordinate method for multi-objective behaviors of the robot. This paper is organized as follows. Section 2 describes our developed robot hardware and control architecture. A motion coordinate method is proposed as a basic control architecture of the robot. Simplified fuzzy inference is used for describing robot’s behaviors. Furthermore, a sen- sory network is applied as the perception mechanism of the robot. Section 3 proposes a behavior learning method for the mobile robot. Section 4 shows experimental results of the behavior coordinate and behavior learning of our devel- oped robot. 2. A Mobile Robot with Fuzzy Controller 2. 1 Hardware architecture of a mobile robot We developed a mobile robot shown in Fig. 1. The diam- eter of the robot is 32.0 [cm]. The 32 bit CPU is built in the robot. Figure 2 shows the sensing system of this robot. The robot has infrared proximity sensors which detect obstacles within 10.0 [cm] in each sensing direction. In addition, the robot has ultra sonic sensors that measure the distance to obstacles between 10.0 and 100.0 [cm]. Consequently, the degrees of danger based on distance is measured by the in- frared proximity and ultra sonic sensors. In addition, CdS sensors are equipped to measure the degree of light sur- rounding the robot in 8 directions. Two stepping motors are used for the actuator. By using these motors, the robot can move forward and backward, and can turn right and left. The stepping motor can be controlled by the signal of a motor output level between 200 and 2000. A wire- less communication is used for bidirectional communica- tion between robot and host computer in order to download programs and to perform tele-operation by human opera- tors. Furthermore, a wireless CCD camera system is built in the middle of the robot to perform tele-operation. The human can operate the robot by watching visual images c 1999 Cyber Scientific Paper No. 1345–269X/99/020079-08