<p>In recent years, there is a strong need for autonomous robots to solve labor shortages and perform tasks in harsh environments. These robots need to avoid obstacles and plan their paths by analyzing data acquired in real environments and generating environmental maps. Unsupervised learning using Growing Neural Gas (GNG) is an algorithm that dynamically constructs a graph structure based on 3D point clouds acquired from sensors. GNG is characterized by its high robustness to downsampling and inhomogeneous data. On the other hand, although clustering based on the constructed topological structure is possible, it is not possible to recognize objects after clustering. To perform a wide range of tasks, the ability to recognize objects is essential, and deep learning methods are effective for this purpose. In particular, PointNet was the first point cloud deep learning model to deal with the irregularity and order invariance of point clouds. In this study, we propose a new deep learning model, PointGNG, which integrates GNG with PointNet++ and extracts local structural features using adjacency-based neighborhood information. Using this method, the proposed method achieved accuracy comparable to PointNet++ on the ModelNet40 benchmark. A slight improvement in accuracy was also observed for other datasets that include background information. This study demonstrates that the integration of unsupervised learning and deep learning with GNG is effective for efficient processing of point cloud data and object recognition. Furthermore, we evaluated the robustness of the proposed model and examined how the choice of training dataset affects its adaptability to real-world environments.</p>