Intention Inference and Decision Making with Hierarchical Gaussian Process Dynamics Model

Anticipation is crucial for fluent human-robot interaction, which allows a robot to independently coordinate its actions with human beings in joint activities. An anticipatory robot relies on a predictive model of its human partners, and selects its own action according to the model’s predictions. Intention inference and decision making are key elements towards such anticipatory robots. In this thesis, we present a machine-learning approach to intention inference and decision making, based on Hierarchical Gaussian Process Dynamics Model (H-GPDM). We first introduce the H-GPDM, a class of generic latent-variable dynamics models. The HGPDM represents the generative process of complex human movements that are directed by exogenous driving factors. Incorporating the exogenous variables in the dynamics model, the HGPDM achieves improved interpretation, analysis, and prediction of human movements. While exact inference of the exogenous variables and the latent states is intractable, we introduce an approximate method using variational Bayesian inference, and demonstrate the merits of the H-GPDM in three different applications of human movement analysis. The H-GPDM lays a foundation for the following studies on intention inference and decision making. Intention inference is an essential step towards anticipatory robots. For this purpose, we consider a special case of the H-GPDM, the Intention-Driven Dynamics Model (IDDM), which considers the human partners’ intention as exogenous driving factors. The IDDM is applicable to intention inference from observed movements using Bayes’ theorem, where the latent state variables are marginalized out. As most robotics applications are subject to real-time constraints, we introduce an efficient online algorithm that allows for real-time intention inference. We show that the IDDM achieved state-of-the-art performance in intention inference using two humanrobot interaction scenarios, i.e., target prediction for robot table tennis and action recognition for interactive robots. Decision making based on a time series of predictions allows a robot to be proactive in its action selection, which involves a trade-off between the accuracy and confidence of the prediction and the time for executing a selected action. To address the problem of action selection and optimal timing for initiating the movement, we formulate the anticipatory action selection using Partially Observable Markov Decision Process, where the H-GPDM is adopted to update belief state and to estimate transition model. We present two approaches to policy learning and decision making, and show their effectiveness using human-robot table tennis. In addition, we consider decision making solely based on the preference of the human partners, where observations are not sufficient for reliable intention inference. We formulate it as a repeated game and present a learning approach to safe strategies that exploit the humans’ preferences. The learned strategy enables action selection when reliable intention inference is not available due to insufficient observation, e.g., for a robot to return served balls from a human table tennis player. In this thesis, we use human-robot table tennis as a running example, where a key bottleneck is the limited amount of time for executing a hitting movement. Movement initiation usually requires an early decision on the type of action, such as a forehand or backhand hitting movement, at least 80 ms before the opponent has hit the ball. The robot, therefore, needs to be anticipatory and proactive of the opponent’s intended target. Using the proposed methods, the robot