Planning for Human-Robot Teaming Kartik Talamadupula † and Subbarao Kambhampati † and Paul Schermerhorn § and J. Benton † and Matthias Scheutz § † Department of Computer Science Arizona State University Tempe, AZ 85287 USA {krt,rao, j.benton} @ asu.edu § Cognitive Science Program Indiana University Bloomington, IN 47406 USA {pscherme,mscheutz} @ indiana.edu Abstract One of the most important applications of planning technol- ogy has been – and continues to be – guiding robotic agents in an autonomous fashion through complex problem scenarios. Increasingly, real-world scenarios are evolving in a way that includes humans as actors in the loop along with the robot and the planning system. These humans are stakeholders whose roles may vary between that of a commander or a system or domain expert; the one common thread is that together with the robot, they form a team that shares common goals. In this paper, we consider challenges posed by such human- robot teaming scenarios from a purely planning-centric per- spective, and discuss the dimensions of variation within ap- plication problems in such scenarios. We seek to differenti- ate planning for human-robot teaming from the general area of human-robot interaction, since we are mainly interested in the planning tools that facilitate such teaming. We look at some problems that are encountered in deploying existing planning techniques in such teaming scenarios, and illustrate these with our experience in a real world search and rescue scenario. We follow this up with results from runs involving a robot controlled by a planner whose goal handling capabil- ities are augmented. Introduction One of the earliest motivations for Artificial Intelligence as a field of study was to provide autonomous control to robotic agents that carry out useful service tasks. Application sce- narios for these kinds of tasks span a wide spectrum that includes military drones and mules, household assistance agents (Goebelbecker et al. 2010) and search and rescue robots (Schermerhorn et al. 2009). The concept of team- ing between humans and robots is central to all these ap- plications – the notion of robotic agents that support a hu- man agent’s goals while executing autonomously is a recur- ring theme. The level of autonomy that is desired of these robotic agents is often achievable only by integrating them with planning systems that can not only plan for the initially specified goals, but also updates to these goals as well as changes to the world and to the agent’s capabilities. Recent years have seen the emergence of fast planning algorithms and systems that can account for a large num- ber of the features that distinguish a real world application from a theoretical scenario – time, cost, resources, uncer- tainty and execution failure. Though planners of the past have been able to model many of these features (Penberthy and Weld 1995), the scale-up that is required to support real world time windows has only come about in the past decade due to the use of heuristic search methods for plan synthesis. Current planners still operate under a number of restrictive assumptions, and classical planners like LAMA (Richter and Westphal 2010) are clearly the fastest of the lot. The chal- lenge then is one of identifying the features that are essential when considering planning support for such joint human- robot endeavors, and of providing a general framework for these problems. This problem is quite distinct from the ex- isting field of human-robot interaction (HRI), since we are interested more in what existing planning techniques can be used or extended in order to facilitate teaming scenar- ios. Towards this end, we discuss a new class of problems under the collective term human-robot teaming (HRT), and present the essential dimensions of such problems with re- spect to planning. The teaming aspect of these problems arises from the fact that the human and the robot are both acting towards achieving the same set of shared goals, and the relationship between them can be defined in terms of known modes of interactions in teams (e.g. colleagues, commander-subordinate, etc.). Though there has been work in the past on the intersection of tasks involving humans, robots and planners, most of that work has concentrated on a system-centric view of the interaction. Our focus in this paper is instead on the teaming, and on describing the char- acteristics of this problem as applicable to planning. The rest of this paper is organized as follows: we first discuss the concept of human-robot teaming and list the di- mensions of interest to planning in such scenarios. Follow- ing this, we look at a search and rescue scenario that we had to provide planning support for as a case study, and place it within the HRT spectrum. We then detail some initial work that we have undertaken in tackling some planning challenges inherent in teaming scenarios. We discuss two specific problems – handling incomplete models, and the problem of goal specification and revision. As part of the latter, we also detail our work with a robot executing in a real-world search and rescue scenario, and present the ag- gregated results of the robot’s runs through this task guided by our planning system. Our hope is for this paper to serve as a catalyst that spurs the planning community into further defining and mapping the application-rich field of human-