Emotional body language synthesis for humanoid robots

Abstract

In the next decade, societies will witness a rise in service robots deployed in social environments, such as schools, homes, or shops, where they will operate as assistants, public relation agents, or companions. People are expected to willingly engage and collaborate with these robots to accomplish positive outcomes. To facilitate collaboration, robots need to comply with the behavioural and social norms used by humans in their daily interactions. One such behavioural norm is the expression of emotion through body language.

Previous work on emotional body language synthesis for humanoid robots has been mainly focused on hand-coded design methods, often employing features extracted from human body language. However, the hand-coded design is cumbersome and results in a limited number of expressions with low variability. This limitation can be at the expense of user engagement since the robotic behaviours will appear repetitive and predictable, especially in long-term interaction. Furthermore, design approaches strictly based on human emotional body language might not transfer effectively on robots because of their simpler morphology. Finally, most previous work is using six or fewer basic emotion categories in the design and the evaluation phase of emotional expressions. This approach might result in lossy compression of the granularity in emotion expression.

The current thesis presents a methodology for developing a complete framework of emotional body language generation for a humanoid robot, intending to address these three limitations. Our starting point is a small set of animations designed by professional animators with the robot morphology in mind. We conducted an initial user study to acquire reliable dimensional labels of valence and arousal for each animation. In the next step, we used the motion sequences from these animations to train a Variational Autoencoder, a deep learning model, to generate numerous new animations in an unsupervised setting. Finally, we extended the model to condition the generative process with valence and arousal attributes, and we conducted a user study to evaluate the interpretability of the animations in terms of valence, arousal, and dominance. The results indicate moderate to strong interpretability.