Improving fidelity of close social interaction animations in social VR with a machine learning-based refinement framework

Social Virtual Reality platforms enable users to embody avatars and interact in virtual worlds. While research suggests that full-body avatar representations are generally preferred, high behavioral fidelity in avatar animations can be hard to achieve. Hardware-based tracking can be particularly effective but is costly, whereas Inverse Kinematics (IK) is more affordable but less accurate, leading to less realistic motion. Recent neural network-based approaches have shown promise in improving IK-based animations by predicting natural movements; however, guaranteeing high levels of fidelity in avatar-to-avatar interactions, particularly those involving close contact, remains challenging even with those approaches. With the aim to address such issue, this paper proposes a neural network-based refinement framework to enhance behavioral fidelity in close social interactions. To investigate its effectiveness, hugging has been selected as a use case. The framework, trained on motion capture data, has been evaluated via a user study, showing improved behavioral fidelity in avatar social interactions.