Planning for Human-Robot Interaction:Representing Time And Human Intention

Abstract

This thesis proposes a novel approach to planning for a specific class of human-robot interaction domains: those in which robots engage in tasks with humans that are governed by social conventions. When humans perform these social tasks, they try to achieve their own goals in an environment that they share with other people. Each participant has their own individual goals, and the interaction is neither purely cooperative nor adversarial in nature. However, the actions of others may have a direct impact on each person’s ability to achieve their goals.Social conventions exist as a guideline for how to interact with others so that all parties involved can achieve their goals efficiently without interfering with one another. Recognizing what goals others are trying to achieve and performing actions at the appropriate time in the interaction are critical abilities for social competence. Adding a robot into these systems without upsetting the social equilibrium is challenging. The approach to this problem taken in this thesis focuses on creating more accurate models of social tasks for planning. Because the human participants are modeled as a part of the environment, the world state in these problems is dynamic and partially observable. Specifically, the intentions of the humans are represented as hidden state in a partially observable Markov decision process (POMDP), and the time-dependence of action outcomes are explicitly modeled for both the humans and the robot. A model structure designed by a human expert is combined with experimental data of humans performing the task in question. The resulting models are large and complex. State aggregation over the time dimension of the state space is used to trade off between the quality of the representation of time and the model’s size in order to find sufficiently expressive models that can also be solved tractably. The performance of the policies obtained are evaluated both in simulation and in interactions with human participants. The utility of this approach is demonstrated by using it to implement a controller for a mobile robot that rides elevators with people and an agent in a driving simulator that attempts to take the Pittsburgh left at an intersection with human drivers. Performance is evaluated by comparing the policies obtained using the proposed modeling technique to policies developed using less expressive representations, and by evaluating objective performance criteria and people’s subjective responses to interacting with the agent in the simulated driving task. The policies for time-dependent POMDP models with human intention as hidden state outperformed the policies of the less expressive models, achieving both higher rewards during interaction and more positive evaluations for naturalness and social propriety of behavior.