Interactive Task Planning through Multiple Abstraction: Application to Assistant Robotics Cipriano Galindo and Javier Gonzalez and Juan-Antonio Fernandez-Madrigal 1 1 INTRODUCTION Assistant robotics has become an emergent field within the robotic and artificial intelligence communities ([1],[2]). The main character- istic of assistant robots is that they are designed to serve non-expert people within their environment. They must plan and act efficiently to accomplish tasks specified in a human-like manner, i.e. "take this envelope to Peter’s office". Thus, an assistant robot must manage a symbolic model of its environment -its world model- that involves human concepts. A human-inspired world representation can be used to endow as- sistant robots with that mentioned capability. It is stated in literature that humans widely use the mechanism of abstraction ([3],[6],[8]). Abstraction allows humans to work with abstract concepts that group other, more particular concepts. For instance, when somebody refers to "my office" she/he is talking about an abstract concept, dropping unnecessary details like door, wall, table, chair, cabinet, etc. A robot managing a world model that includes human concepts allows users to specify tasks in a human-like manner. Furthermore, the user can also interact with the robot during the planning process since the robot reports an understandable plan to the human. The work presented in this paper is intended to jointly cover hu- man interaction and task planning efficiency, by using a hierarchical model of the environment. For our purposes, a multihierarchical world model called Multi-AH-graph ([6],[7]) has been implemented with two hierarchies of abstraction: the task planning hierarchy de- voted to efficient task planning and the cognitive hierarchy engaged in human communication. The use of these two hierarchies requires a translation process to transform concepts from one hierarchy to the other, which is addressed in this paper. The paper is structured as follows. Section 2 briefly reviews the Multi-AH-graph model. Section 3 is devoted to describe the human interaction mechanism in robot task planning and its application to a real robotic application. Finally, some conclusions and future work are outlined. 2 THE MULTI-AH-GRAPH MODEL A Multi-AH-graph is a graph representation that includes hierarchical information possibly in more than one hierarchy. Ab- straction produces different layers, called hierarchical levels, that can represent symbolically the same environment with different amount of detail. These layers are flat graphs whose nodes represent elements 1 Work supported by the Spanish Government, DPI2002-01319. The authors are with the System Engineering and Automation Department, Univer- sity of Malaga, Campus Teatinos - Complejo Tecnológico, 29071, Málaga, Spain. E-mail:{cipriano,jgonzalez,jafma}@ctima.uma.es of the environment (concepts), and whose arcs represent different re- lations between those concepts (for example, "navigability"). A sim- ple example of a Multi-AH-graph with one hierarchy that models a typical scenario is shown in fig. 1. Distinctive Place a) b) 6 16 9 10 8 4 18 5 15 11 17 3 2 14 7 13 12 1 Laboratory Office 1 Office 2 Library Northern Rooms Southern Rooms Environment Laboratory Office 1 Corridor Corridor Rooms Library Office 2 Northern Rooms Corridor Rooms Southern Rooms c) 6 16 9 10 8 4 18 5 15 11 17 3 2 14 7 13 12 1 Corridor Figure 1. An example of an Multi-AH-graph (with one hierarchy). (a) A schematic plan of a real environment. (b) The lowest hierarchical level (ground level). (c) Upper levels of the hierarchy. Broadly speaking, a Multi-AH-graph is a set of hierarchies in- terwoven in a directed acyclic graph structure, where each hierar- chy level may can be shared by others hierarchies. Using a sym- bolic multihierarchical representation of the environment has demon- strated to provide important benefits for robot operation (please refer to [6], [7] for more detail). 3 HUMAN INTERACTION IN ROBOT TASK PLANNING Among other advantages [7], the Multi-AH-graph model provides efficiency in hierarchical task planning [4]. In this paper we use a multihierarchical scheme to enable a human to supervise and inter- act with the robot task planner. In our scheme, a translation process (section 3.1) is required to shift concepts between two hierarchies managed by the robot.