This work presents an approach to vision-based robotic grasping, combining scene understanding with advanced neural network architectures. The primary goal is to develop a system that enables robots to grasp a wide variety of objects in unstructured environments using RGB-D data. The methodology integrates two key components: the YOLOv8 model for real-time object detection and the Generative Residual Convolutional Neural Network (GRCNN) for precise grasp pose inference. The experimental setup employs the PyBullet simulation environment, utilizing a Universal Robots UR5 model equipped with a Robotiq 2F-85 parallel-jaw gripper. The YOLOv8 model is trained using a custom dataset created in PyBullet and annotated via the Roboflow platform, ensuring accurate object detection and localization. The GRCNN model generates grasp quality maps, including quality, angle, and width maps, which are crucial for determining the optimal grasp points and orientations. The system is thoroughly tested in the PyBullet environment, where it demonstrates robust performance across various scenarios, accurately detecting objects and planning grasps in real-time. The simulation results are validated through real-world experiments using the UR5 robot, confirming the system’s effectiveness in practical applications. This work also addresses the limitations of traditional grasping methods, which often rely on pre-programmed instructions and detailed 3D models, making them unsuitable for dynamic environments. By leveraging deep learning and real-time processing, the proposed system overcomes these challenges, providing a scalable and efficient solution for robotic grasping in unstructured settings. Future work will involve optimizing neural networks, exploring adaptive grasping strategies, and enhancing language understanding capabilities to improve system responsiveness and versatility in complex tasks.