I See What You Did There: Towards a Gaze Mechanism for Joint Actions in Human-Robot Interaction

Abstract

We imagine that service robots must collaborate with humans in physical object manipulation tasks to be of assistance in everyday scenarios, such as setting a table. This collaboration requires the capability of joint attention to smoothly accomplish a shared goal. One special modality for joint attention is the gaze behavior of an actor. Herein, we discuss the human gaze in physical tasks and its underlying cognitive mechanisms, a novel probabilistic robotic gaze controller in object-centred collaborative physical tasks, and its inclusion in a well-known joint action human-robot interaction (HRI) benchmark. First, we discuss human gaze behavior as an important modality for signaling, detecting, and monitoring joint attention processes. This is followed by an overview of joint attention implementations in HRI and commonly used artificial intelligence methods for planning and plan recognition. These methods are used to mimic qualities of different components in psychological joint attention models in humans. In object manipulation tasks, the gaze behavior is not only used to gather information about the environment, but also has a communicative role, as the gaze direction can be interpreted by the interaction partner. The intended actions and beliefs about the current world state are communicated through the gaze. We argue that robotic gaze behavior, which humans easily interpret, will improve the interaction capability of a social robot. We investigate this claim in an already established HRI joint action benchmark scenario of collaboratively building a tower out of different blocks. To this end, we propose a stochastic gaze controller for joint action tasks and present results of a pilot study.