A Reinforcement Learning Based Cognitive Empathy Framework for Social Robots
- PDF / 1,277,107 Bytes
- 15 Pages / 595.276 x 790.866 pts Page_size
- 58 Downloads / 253 Views
A Reinforcement Learning Based Cognitive Empathy Framework for Social Robots Elahe Bagheri1
· Oliver Roesler2 · Hoang-Long Cao1 · Bram Vanderborght1
Accepted: 19 July 2020 © Springer Nature B.V. 2020
Abstract Robots that express human’s social norms, like empathy, are perceived as more friendly, understanding, and caring. However, appropriate human-like empathic behaviors cannot be defined in advance, instead, they must be learned through daily interaction with humans in different situations. Additionally, to learn and apply the correct behaviors, robots must be able to perceive and understand the affective states of humans. This study presents a framework to enable cognitive empathy in social robots, which uses facial emotion recognition to perceive and understand the affective states of human users. The perceived affective state is then provided to a reinforcement learning model to enable a robot to learn the most appropriate empathic behaviors for different states. The proposed framework has been evaluated through an experiment between 28 individual humans and the humanoid robot Pepper. The results show that by applying empathic behaviors selected by the employed learning model, the robot is able to provide participants comfort and confidence and help them enjoy and feel better. Keywords Empathy · Reinforcement learning · Personality · Human–robot interaction · Social robot
1 Introduction The number of robots interacting with humans has increased and some of these robots are interacting with sensitive groups of society, like users with disabilities [8], elders [5,21], people with dementia [9], and children with autism [12], to provide instructions, feedback, and support. Previous studies showed humans prefer robots that show human’s social norms, like empathy, which makes robots appear more caring, supportive [7], engageable [23], and friendly [33]. In addition, robots have the opportunity to serve as empathizers in circumstances that humans tend to avoid empathy, e.g., if helping prevents obtaining a desired outcome [11], is exhausting [35] or causes cost [39]. Even though the positive effects of empathy in humanrobot interaction (HRI) have been established, a coherent empathic model has not been obtained yet [22], which is mainly due to two reasons. The first reason is that most pro-
B
Elahe Bagheri [email protected]
1
Robotics and Multibody Mechanics Research Group, Vrije Universiteit Brussel and Flanders Make, Brussels, Belgium
2
Artificial Intelligence Lab, Vrije Universiteit Brussel, Brussels, Belgium
posed empathic models consider contextual information and task-related factors, e.g., the user’s status in playing a game, i.e., failure or success [25], to conduct empathy, while user’s affective state is ignored. In contrast, the model proposed by Churmani et al. [10] considered user’s affective state and to determine it asked the users about their feeling through a questionnaire. Whilst the latter approach takes the affective state into account, assuming that users provide accurate informatio
Data Loading...
