In this project, we aim to develop and evaluate a multi-modal AI agent capable of interacting with and manipulating 3D environments using Blender as a simulation platform. The agent will be built upon large vision-language models (e.g., Qwen3-VL) and trained through a combination of supervised fine-tuning (SFT), reinforcement learning (RL) to interpret multi-modal instructions, perceive complex 3D scenes, and perform tasks such as object manipulation, scene editing, and dynamic camera control. By integrating perception, vision-language understanding, and action execution within a unified framework, this work advances the frontier of embodied AI in creative 3D domains, with potential applications in virtual production, educational tools, and generative 3D content creation.