Reasoning about humans and its use in a cognitive control architecture for a collaborative robot Rachid Alami, Aurélie Clodic CNRS, LAAS, 7 avenue du colonel Roche, F 31400 Toulouse, France Univ de Toulouse, LAAS, F 31400 Toulouse, France rachid.alami@laas.fr, aurelie.clodic@laas.fr Raja Chatila Institut des Systèmes Intelligents et de Robotique Université Pierre et Marie Curie-Paris6, CNRS UMR 7222 4, Place Jussieu 75252 Paris Cedex 05 - France raja.chatila@isir.upmc.fr Séverin Lemaignan Computer-Human Interaction in Learning and Instruction Laboratory (CHILI) Ecole Polytechnique Fédérale de Lausanne (EPFL) CH-1015 Lausanne, Switzerland severin.lemaignan@epfl.ch Keywords Human-robot collaboration, Affordances, Human-aware task planning 1. INTRODUCTION We discuss here a decisional framework for human-robot in- teractive task achievement that is aimed to allow the robot not only to accomplish its tasks but also to produce behav- iors that support its engagement vis-a-vis its human partner and to interpret human behaviors and intentions. We have adopted a constructive approach based on effective individ- ual and collaborative skills. The system is comprehensive since it aims at dealing with a complete set of abilities ar- ticulated so that the robot controller is effectively able to conduct a collaborative task with a human partner in a flex- ible manner These abilities include geometric reasoning and situation assessment based essentially on perspective-taking and affordances, management and exploitation by the robot of each agent beliefs (human and robot) in a separate cog- nitive model, human-aware task planning and human and robot interleaved plan achievement. 2. FRAMEWORK Figure 1 illustrates the main concepts on which the system is built. Collaboration happens as a consequence of an explicit request of the human to satisfy a goal or because the robot finds itself in a situation where it is useful if not manda- tory. In both cases, the robot has a goal to satisfy. An important issue is the notion of engagement, a process in which the robot will have to establish, maintain and termi- nate a connection with a human partner. This covers goal establishment, selection of an incremental refinement of the task that is intended to be achieved, and execution control including monitoring, and even influencing, human task per- ? ? Hello! Multi-modal Dialog Mutual Activity Observation Figure 1: The conceptual framework: Human and Robot share task and space. They are in a mutual observation situation. Robot explicity reasons about the fact that it is (or has to be) perceived by its human partner. Planning is performed at high-level symbolic as well as as geometric level formance and his/her commitment to the goal. The human involvement may range from a direct participation to the task achievement, to a simple “acceptance” of robot activity in his/her close vicinity. Our robot is controlled by a three layer architecture [1]. We describe in the sequel the main activities performed by the robot controller (Figure 2.): (1) Situation assessment and context management (2) Goals and plans management (3) Action refinement, execution and monitoring. Other deci- sional activities, such as situated dialog ([9, 4], not presented here) have been developed that use the same set of compo- nents. 2.1 Situation Assessment and Context Management Geometric reasoning plays a central role in our architec- ture. It is performed by a component called SPARK (Spa- tial Reasoning and Knowledge [12]). It is responsible for geometric information gathering and it embeds a number of decisional activities linked to abstraction (symbolic facts