Creating a real-world fruit tree dataset is notoriously time-consuming and expensive, mainly due to the challenges in data collection and labeling. In contrast, the proposed method generates a synthetic dataset that is inherently labeled and can be produced more rapidly and cost-effectively. This synthetic environment plays a crucial role in generating comprehensive data for deep learning applications, particularly for enhancing pre-trained deep learning models through reinforcement learning. By utilizing synthetic datasets, we facilitate the fine-tuning of pre-trained deep learning models. This method not only improves the capability and efficiency of AI models in real orchard scenarios but also offers a scalable and cost-effective solution for training complex models. The implications of this research are significant, suggesting a future where AI-driven robots can be trained more effectively and swiftly for a variety of agricultural tasks, leading to increased productivity, reduced labor costs, and sustainable farming practices. Ongoing studies are specifically focusing on automatic label creation for different deep learning algorithms, such as YOLO (You-Only-Look-Once), and various 3D rendering frameworks. This development aims to automate and optimize the labeling process, ensuring that datasets are not only comprehensive but also tailored to the specific requirements of different deep learning frameworks.