Background: Emotion recognition skills are predicted to be fundamental features in social robots. Since facial detection and recognition algorithms are compute-intensive operations, it needs to identify methods that can parallelize the algorithmic operations for large-scale information exchange in real time. The study aims were to identify if traditional machine learning algorithms could be used to assess every user emotions separately, to relate emotion recognizing in two robotic modalities: static or motion robot, and to evaluate the acceptability and usability of assistive robot from an end-user point of view.
Methods: Twenty-seven hospital employees (M = 12; F = 15) were recruited to perform the experiment showing 60 positive, negative, or neutral images selected in the International Affective Picture System (IAPS) database. The experiment was performed with the Pepper robot. Concerning experimental phase with Pepper in active mode, a concordant mimicry was programmed based on types of images (positive, negative, and neutral). During the experimentation, the images were shown by a tablet on robot chest and a web interface lasting 7 s for each slide. For each image, the participants were asked to perform a subjective assessment of the perceived emotional experience using the Self-Assessment Manikin (SAM). After participants used robotic solution, Almere model questionnaire (AMQ) and system usability scale (SUS) were administered to assess acceptability, usability, and functionality of robotic solution. Analysis wasperformed on video recordings. The evaluation of three types of attitude (positive, negative, andneutral) wasperformed through two classification algorithms of machine learning: k-nearest neighbors (KNN) and random forest (RF).
Results: According to the analysis of emotions performed on the recorded videos, RF algorithm performance wasbetter in terms of accuracy (mean ± sd = 0.98 ± 0.01) and execution time (mean ± sd = 5.73 ± 0.86 s) than KNN algorithm. By RF algorithm, all neutral, positive and negative attitudes had an equal and high precision (mean = 0.98) and F-measure (mean = 0.98). Most of the participants confirmed a high level of usability and acceptability of the robotic solution.
Conclusions: RF algorithm performance was better in terms of accuracy and execution time than KNN algorithm. The robot was not a disturbing factor in the arousal of emotions.
Keywords: acceptability; human-robot interaction; monitoring of behaviorand internal states of humans; non-verbal cues and expressiveness.