Continual Learning for Adaptive Affective Human-Robot Interaction

• Published in: 2022 10th International Conference on Affective Computing and Intelligent Interaction Workshops and Demos (ACIIW) pp. 1 - 5
• Main Author: Maharjan, Rahul Singh
• Format: Conference Proceeding
• Language: English
• Published: IEEE 18.10.2022
• Subjects: Adaptation models; Affective computing; Biological system modeling; Computational modeling; Continual Learning; Deep learning; Emotion recognition; Human-Robot Interaction; Service robots; Social Robotics
• DOI: 10.1109/ACIIW57231.2022.10086015

Unlike industrial robots, social robots such as Pepper, Nao, Furhat, and Moxi are being developed with the goal of assistance, collaboration, and even companionship to humans. However, such social robots at the moment are not suitable for continually changing real-world due to the inability to understand and adapt to the social ecosystem of human beings. Driven by advancements in affective computing and deep learning, social robots aim to recognise affective, social clues such as emotion for interaction. Existing deep learning based affective computing models for human-robot interaction are based on static data distribution; however, in the ever-changing real-world environment, the static nature of data distributions gets invalid. As a consequence, the performance of such social robots degrades substantially. This paper argues that continual learning based affective computing is a step forward for adaptive and personalisable human-robot interaction in a continually changing real world. We conducted an initial experiment on how a continual learning approach can be integrated into emotion recognition and observed that the continual learning approach could learn new emotions in a continuous setting. The intention is to use this insight to build more complex continual emotion recognition models that can adapt and personalise for support and collaboration in the social ecosystem.