Virtual avatars have gained increasing attention in Extended Reality (XR) research because of their potential to support social presence, communication, and engagement in immersive environments, where scene-aware and believable avatar behavior is desirable. This thesis investigates how text-generated human motion can be grounded into executable avatar actions in a Unity environment representing XR scenarios. To address this problem, the work integrates three separate modules consisting of intent reasoning with a Large Language Model (LLM), 3D affordance grounding, and text-to-motion generation.

The modular frontend-backend pipeline captures object semantics, point clouds, and avatar state in Unity, processes these inputs on a backend server, and returns generated motion for retargeted avatar execution. The system was evaluated on sitting and lying interactions with chairs, benches, and beds using latency measurements, Root Mean Square Error (RMSE), semantic action quality, and perceived motion quality.

The results show that the pipeline can translate scene data into executable avatar actions within an interactive response range, while remaining contact artifacts indicate that single-target grounding is not sufficient for fully natural avatar-object interaction. This study demonstrates both the feasibility and limitations of combining LLM reasoning, affordance detection, and text-to-motion generation for grounded avatar interaction. The findings highlight the need for stronger object-conditioned motion generation, explicit user intent, and richer interaction constraints to improve robustness in future XR avatar systems.