Modeling Interaction Structure for Robot Imitation Learning of Human Social Behavior

Abstract

This study presents a learning-by-imitation technique that learns social robot interaction behaviors from natural humanhuman interaction data and requires minimum input from a designer. To solve the problem of responding to ambiguous human actions, a novel topic clustering algorithm based on action cooccurrence frequencies is introduced. The system learns humanreadable rules that dictate which action the robot should take, based on the most recent human action and the current estimated topic of conversation. The technique is demonstrated in a scenario where the robot learns to play the role of a travel agent. The proposed technique outperformed several baseline techniques in qualitative and quantitative evaluations. It responded more accurately to ambiguous questions and participants found it was easier to understand, provided more information, and required less effort to interact with.

Neural-network-based Memory for a Social Robot

“…In an ideal case, the number of GRU units that learn to encode memories should match the number of memory types that occurred in the training dataset (10). However, since the utterances in memory-setting actions have many natural variations of phrasing, it is challenging for the neural network to learn a perfect memory representation.…”

“…More recently, Doering et al introduced a data-driven behavior learning system for an android travel agent that learns interaction logic in the form of human-readable interaction rules [10]. In contrast to the work by Liu et al, the system contains a model of the time-persistent topic of conversation that helps to resolve ambiguous customer questions.…”

“…Researchers have recently explored the role of social robots in a variety of domains, including elder care [6,7], shop-keeping [1,2,4,3,55], hotels [8], education [9], travel [10], and as personal 24:2 M. Doering et al companions [11,12]. As the capability of artificial intelligence increases, this trend is likely to continue.…”

“…In response to this challenge, previous works have proposed data-driven approaches for designing social robot behaviors automatically, with minimal input from a human designer, by imitation learning from examples of natural human-human interaction [1,2,4,3,10,55]. While impressive, with these systems the robot is limited to responding to contextual information from within only a short temporal window of the overall interaction.…”

Autonomously Learning One-To-Many Social Interaction Logic from Human-Human Interaction Data

“…While much research has been devoted to learning interaction logic from human-human data for one-to-one scenarios [7,20,21], one-to-many scenarios introduce novel challenges. Consider Fig.…”