Robot Task Learning and CollaborationGuy HoffmanMIT Media [email protected] LockerdMIT Media [email protected] this paper we describe our work towards human-robotcollaborative learning of tasks and the subsequent cooper-ative execution framework. In our work, a humanoid robotlearns hierarchical tasks comprised of primitive action andsmaller tasks, and is able to perform this task joinly with ahuman partner. Both the tutoring and the execution of thetasks are viewed as a collaborative discourse, building onnatural human social skills and conventions.Our approach is a goal-centric one, using goals at boththe task and the action level to establish common ground inlearning and collaboration. During the execution stage, dy-namic meshing of subplans and self-assessment provide aturn-taking mechanism based on mutual commitment andsupport, resulting in a shared collaborative activity that isintuitive for a human participant.1. IntroductionIn the Robotic Life group at the MIT Media Lab weare working towards building social machines that are nat-ural and intuitive for people to interact with. We wouldlike to enable machines to take advantage of the multitudeof social skills that humans exhibit when interacting witheach other: communication, cooperation, social learning.This paper details our work towards supplying our robot,Leonardo, with the ability to learn tasks through interac-tion with a human teacher and then collaboratively executethese tasks with a human partner.2. ApproachOur aim is to teach the robot a structurally complex taskto later be performed collaboratively with a human. Thisis made possible through a goal oriented representation oftasks that affords the construction of joint intentions. Thissection provides an overview of the three main componentsof our system: hierarchical learning, goals, and joint inten-tions.2.1. Hierarchical Task LearningIn our implementation the robot learns a representationof a new task, its constituent actions and sub-tasks, as wellas goals associated with each of these. The task represen-tation is such that it then affords the construction of sharedplans and joint intentions in order to complete the task witha partner.A number of social and expressive skills contribute tothe robot’s effectiveness in understanding and collaborat-ing on a complex task with the human teacher. The tutor-ing of tasks exemplifies our approach to teaching as a col-laborative discussion. Join attention is established both onthe object level and on the task structure level. Leonardouses subtle expression to indicate to the human tutor whenhe is ready to learn something new, and his performanceof taught actions provides the tutor with immediate feed-back about the robot’s comprehension of the task. Enve-lope displays such as gaze aversion, eye contact and sub-tle nods and are used to segment a complex task learningstructure in a natural way to the tutor. Natural key wordssuch as ”next”, ”first” are used to indicate task structure andsequencing constraints [5].2.2. Goal Driven ActionWe believe that a goal-centric view is a fundamental fea-ture of both teaching and collaboration. It has been repeat-edly shown that humans interpret intentions based on goals[11, 7, 1] and that goals, not specific activities or motiontrajectories, are what is most important in collaborative dis-course. Goals provide a common ground for communica-tion and interaction. This is particularly important in thecollaborative setting, since the human partner is biased touse an intention-based psychology to interpret the agent’sactions [6].In the learning of a task, a goal is associated with each ofthe constituent actions as well as the task as a whole. There-fore, the task goal is more than just the conjunction of thegoals of its actions and sub-tasks. Additionally, in execut-ing the task, the task goal can be evaluated as a whole ratherthan evaluating each of its children’s goals to determine ifthe task is done, improving efficiency. We found the goaldriven approach crucial for both the tutoring of tasks andcollaborating on them. Goals provide a common ground foraction segmentation and understanding as well as for coor-dinating actions as part of a collaborative activity.2.3. Joint IntentionIn a collaborative task, a number of agents work togetherto solve a common problem. For this to take place, a jointcourse of action must emerge from the collection of the in-dividual actions of the agents. In human collaboration, thisdoes not reduce to just the sum of the individual actions de-rived from individual intentions, but rather is an interplayof actions inspired and affected by a joint or group inten-tion.Several models have been proposed to explain how jointintention results in individual intention and action to form ajoint action. Searle [10] claims that collective intent and ac-tion cannot be formalized as a function of the individual in-tentions of the agents involved, but rather that the individualintentions are derived from their role in the common goal.He also stresses the importance of a social infrastructure tosupport collaboration. Bratman [2] breaks down Shared Co-operative Activity into mutual responsiveness, commitmentto the joint activity and commitment to mutual support. Healso introduces the idea of meshing subplans, which ourproject generalizes to dynamically meshing subplans. Co-hen et al [4, 9] claim that a robust collaboration schemein a changing environment with partial knowledge and be-liefs requires communication, commitment to the joint task,commitment to mutual support, and dynamic meshing ofsubplans and action steps.Our implementation tests some of the theoretical claimsregarding joint intention. In the spirit of Bratman’s SCA, weplaced a high importance on communicating the robot’s per-ceived state of the world and the task.A world view centered on goals is important both forthe ability to view a joint action with respect to a particu-lar goal, and for the definition of individual intents based onsub-goals of the common intention. Our goal oriented taskrepresentation affords task collaboration between the robotand a human partner. Goals refer to both world and activ-ity state, establishing common ground between the robotand the human. As a result, joint intention, attention andplanning is naturally achieved. Throughout the collabora-tion, the human partner has a clear idea as to Leo’s currentsingular intent as part