Planning for Symbiotic Action: (Doctoral Consortium)

The field of Artificial Intelligence (AI) has become extremely prominent in recent times, with the integration of different intelligent components into devices and services we use in everyday life. As the capabilities of such systems become more and more complex, one branch of AI that becomes relevant is that of automated planning or sequential decision making, in order for these components to participate in diverse long term tasks. A key aspect of such systems is increased interaction with humans. Challenges in Human-in-the-Loop Planning (HILP) Classical planning has traditionally emphasized on the efficiency or accuracy of the plan generation process. However, in real world applications, especially involving humans, planners must deal with typical challenges including uncertainty and partial knowledge, and issues involving priorities and authority. Technologies that become crucial in this context involve abilities to dynamically predict, anticipate and adapt to changing needs while making task plans. My research focuses on such aspects of “human-in-the-loop planning”. Modalities of HILP My Research Focus I have looked at two specific ways in which automated planners may interact with humans. First I will describe how planners can enable different types of autonomous behavior of robots sharing their workspace with humans i.e. interacting with human colleagues. Then I will look at possible roles of automated planners in platforms that involve collaboration with or among human planners. The aim of my thesis is to provide planning technologies for and motivate well-informed and principled design of complex symbiotic man-machine systems of the future. Humans as Colleagues Many of today’s robots built for tasks like household assistance or hotel/office service or security guards, do not operate in teams, i.e. they do not have common goals and commitments with humans sharing the environment, and interactions with such agents depart from traditional notions of proximal human-robot teams. Even though these scenarios require significantly different levels of autonomy from the Appears in: Proceedings of the 15th International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2016), J. Thangarajah, K. Tuyls, C. Jonker, S. Marsella (eds.), May 9–13, 2016, Singapore. Copyright c © 2016, International Foundation for Autonomous Agents and Multiagent Systems (www.ifaamas.org). All rights reserved. robot, the underlying theme of autonomy in such settings involves the robot achieving some sense of independence of purpose in so much as its existence is not just defined by the goals of the humans around it but is rather contingent on tasks it is supposed to achieve on its own. Thus the robots in a sense become colleagues rather than teammates. This becomes even more prominent when we consider inter-team and intra-team interactions between multiple independent teams in a human-robot cohabited environment. We postulate that interaction with the human cohabitants in such cases should be similar to how we interact with our human colleagues rather than teammates. This provides many interesting possibilities in modeling autonomous behavior in these scenarios the agents must learn modes of passive or stigmergic collaboration. In [1, 2, 3, 4] I model such interactions at three levels of granularity resources, plans and goals and motivate the need for developing appropriate metrics for quantifying performance in such settings. Much of the challenge in developing such behaviors is in modeling the appropriate interaction constraints (I use integer programming formulations for this purpose). In the following discussion I will briefly introduce two such models. Planning with Resource Conflicts In [4], we look at how robots sharing their workspace with humans, and using shared resources, can plan to minimize conflicts on resource usage. We propose a planner that models the intentions of the human colleagues and produces plans that decouple its resource demands with that of the humans’. Note that there is no explicit or direct interaction here between the human and the robot, the interaction is successful inasmuch as the human’s plan was successful. Resource Profiles. We represent information from predicted plans in the form of resource profiles, that enables the robot to reason with how the environment will evolve with time, at different levels of abstraction. This way we compile the complexity of accounting for individual predictions into