Investigating Motion Perception and Physics-based Methods for Body Shape Diversity in Virtual Avatars

Abstract

Creating virtual environments populated with realistic humanoid avatars is a challenging task, especially when the aim is to represent the human diversity found in the real world. One significant factor in this diversity is body shape, which not only impacts the visual appearance of virtual avatars but also influences their physical motion characteristics. The goal of this work was to explore the perception of different body shapes in virtual avatars and investigate methods for incorporating these variations into character animations.

We first investigated whether body shape influences motion perception in virtual avatars, specifically for walking and jogging animations. We conducted a user study in which participants were shown avatars that either matched or differed from the body shape of the motion's original actor. Participants were asked to rate the consistency between the avatar�s body shape and motion. The Results showed that avatars within the same BMI (Body Mass Index) group were perceived as more consistent. We also conducted a case study on the impact of self-representation (users were embodied in a full-body avatar or not) on presence in virtual reality. We found that even minimal representations, such as animated hands, were sufficient for users to feel a sense of presence in head-mounted display systems. As the case study indicated that the avatar body did not significantly impact presence in VR, we instead focussed on the findings from the first study, which clearly demonstrated the importance of body shape in maintaining motion consistency. These results motivated us to explore methods for incorporating body shape diversity into virtual environments.

We explored two distinct methods for generating body shape-aware character animations, mainly focusing on physics-based avatars. The first approach involved a deep reinforcement learning framework, where the task was to animate a rigid body skeleton and mimic the given motion file. We introduced some body shape parameters which modified the task reward for individual characters, thereby incorporating the effect of mass during training. This resulted in altered final motion and variation of motion based on the body shape. While the generated motions aligned with the anthropometric data of human movement, this pipeline was not easily compatible with traditional animation workflows or game engines. This prompted us to investigate some alternative methods.

The second approach was mainly based on adjusting the inertial parameters of physics-driven avatars. We devised a physics-based controller framework that utilises motion, target body weight, and height as inputs to generate retargeted motions. The approach explored whether postural control parameters, i.e., mass and inertial properties like the centre of mass location, can generate realistic motions across different BMIs. This was achieved by controlling limb balance points using a PD control system. To evaluate the quality of the retargeted motions, we then developed a perceptual framework that compared the generated motions against an \textit{average} motion, which had been perceived as most attractive and least distinctive in previous studies.

In addition to individual animations, we explored potential applications of these physics-based avatars in small-size crowd simulations. Specifically, we examined whether body shape affects the detection of motion clones, which are common in crowd scenarios. Contrary to our hypothesis, body shape had no significant effect on motion clone detection. Interestingly, the inclusion of distinct motions caused motion clones to be more easily detectable.

In conclusion, our work explored the role of body shape in the perception and generation of virtual avatar motion. We primarily explored two methodologies for generating body shape-aware animations and developed a perceptual framework for evaluation. While both methods have certain limitations, the significance of body shape in enhancing diversity among virtual avatars cannot be overlooked. Our findings contribute to advancing the representation of diversity in virtual environments and provide practical insights for future avatar animation techniques. Further exploration into factors such as gender, race, or age and their influence on motion perception could deepen our understanding of how best to retarget motion across diverse body types. Our work offers valuable foundations for generating more natural and realistic human motions in virtual settings.