Ragdoll Matching: a non-learned physics-based approach to Humanoid Animation applied to VR Avatars.

The proliferation of Virtual Reality technologies has intensified the quest for enhanced immersion in virtual experiences. In this context, the simulation of the user's body in the form of a credibly animated Avatar, together with the care in ensuring expected and plausible interactions with the virtual environment, are both crucial factors that contribute to preserving a solid sense of presence in the user experience. This study presents a potential approach to achieve a real-time physics-driven humanoid character animation, implemented in the Unity game engine. We designed and developed a solution that permits to control a physics Ragdoll by providing as reference a target rig animated with a kinematic-based technique. By doing so, the animated character is able to perform desired movements while being subjected to the physics engine, allowing collisions and interactions with the virtual environment. Motion Matching was identified as a well suited animation technique, as it satisfies the need for realistic humanoid movements even when applied to VR avatars. The focus during the design process was to maximize versatility and interoperability of the tools developed. The control policy that pilots the physics Ragdoll is based on simple automatic control techniques: compared with other more recent solutions that rely on machine or deep learning techniques, this approach eliminates the need for a training phase for the physics simulation. Our strategy prioritizes modularity and versatility over perfection of results in a well-known and controlled context, while also maintaining the possibility for independent future improvements to each of the simulation modules.