A Dialogue Control Model Based on Ambiguity Evaluation of Users’ Instructions and Stochastic Representation of Experiences Paper: Rb17-6-2456; November 1, 2005 A Dialogue Control Model Based on Ambiguity Evaluation of Users’ Instructions and Stochastic Representation of Experiences Tetsunari Inamura, Masayuki Inaba, and Hirochika Inoue Dept. of Mechano-Informatics, Graduate School of Information Science and Technology, The University of Tokyo 7-3-1 Hongo, Bunkyo-ku, Tokyo 113-0033, Japan E-mail: inamura@jsk.t.u-tokyo.ac.jp [Received February 3, 2005; accepted July 11, 2005] To operate in everyday environments, robots much ac- complish complex tasks following often mbiguous and uncertain instructions, mainly through advanced in- ference or recognition. We focus on an intelligent human-robot interaction framework that reduces the burden of the user. Robots compensate for ambigu- ities by active sensing and dialogue control through questions and suggestions to users. Robots also use experience to reduce the user’s burden. We propose a criterion for ambiguity evaluation of user instructions, stochastic representation of personal experiences and a dialogue control model for accomplishing tasks in complex environments. We demonstrate the feasibil- ity of our proposal in demonstrate experiment where a robot searches for an object in disorganized work space with ambiguous instructions. Keywords: human-robot interaction, Bayesian networks, dialogue planning, user adaptation, experience based be- havior 1. Introduction Research on humanoid robots has a long history, start- ing with the dynamics and control of biped walking and eventually shifting to intelligent behavior in everyday en- vironments supporting human beings. To provide such support, however, robots must under- stand intentions, speech and other human behavior. Many difficulties remain to be overcome before robots can real- ize “intention understanding” or human intention. We propose that robots store experience shared be- tween human and robots using stochastic information pro- cessing [1]. Shared experience plays an important role both in coexistence of human beings and robots and in realization of support by robot. An enormous amount of shared experience is needed for understanding users under uncertain and incomplete conditions, and communication among humans beings. The concept of shared experience has been demon- strated [1] but detailed algorithm for dialogue manage- ment between human beings and robot remain to be clar- ified. We propose how to control and manage dialogue involving complex situations and vague user instructions. We demonstrate the feasibility of the stochastic approach in such ambiguity. 2. Human-Robot Interaction with Shared Ex- periences Physical interaction systems between human beings and machines proposed include the Oxygen Project [2], Smart Rooms [3] and Intelligent Room [4]. The most im- pressive factor of these systems is their ubiquitous sensors and actuators. Satoh et al. proposed the Robotic Room [5] using the concept of ubiquitous environments. The robotic room has many floor pressure sensors, position and button sen- sors on home electric appliances, etc. One feature of the robotic room is the storage of observation results from user behavior to realize smart user-support behavior Archives of human behavior are represented as archives of sensors in the robotic room. Personal everyday behavior is infinitely varied, so it is difficult to estimate and design support behavior in advance. Existent machines are de- signed to satisfy 80% of users using single user model. The storage of behavior between systems and users, i.e., shared experience influences human-machine interaction because shared experience enables the system to decide adequate behavior strategy for each user. Support behavior requires intelligent functions for ubiquitous systems and humanoid systems as a physical agent. Humanoid systems must speak and act using com- munications differing from ubiquitous systems, making shared experience between users and robots more effec- tive in case of humanoids. Intelligent behavior based on shared experience has the following advantages: Acquisition of novel behavior based on observation and imitation of human behavior. Complements of user vague instructions. Realization of adaptive interfacing meeting user’s needs. Journal of Robotics and Mechatronics Vol.17 No.6, 2005 1