In recent years, robotics and advanced fire prevention technologies have significantly changed forest management and offer promising solutions to reduce the risk of forest fires and the associated maintenance costs. Autonomous vehicles equipped with high-precision LiDAR sensors are increasingly being used for dangerous and labor-intensive forestry work such as clearing undergrowth and debris. A key element of these robotic systems is semantic segmentation of 3D LiDAR point cloud data, which enables accurate environment interpretation and object classification that is critical for tasks such as autonomous navigation and fire risk assessment. This chapter focuses on state-of-the-art deep learning techniques for semantic segmentation of point clouds—including convolutional neural networks (CNNs), point-based neural networks, and transformer-based architectures—and evaluates practical strategies for efficient point cloud processing. The focus is on sampling techniques to reduce the complexity of the segmentation process without compromising performance, and on ground elevation estimation to improve performance on terrains with irregular relief, as the performance of semantic segmentation strongly depends on accurate ground elevation estimation. As there is a critical lack of training datasets specifically tailored to forest environments with highly irregular terrain, combining detailed outdoor forest scans with accurate point-based semantic annotations, an overview of currently available datasets is also provided. Finally, some considerations are made to suggest future research directions that could bridge the gap between cutting-edge research and practical application in forest robotics and autonomous forest management, namely the use of self-supervised techniques to explore semi-automatic and automatic labeling techniques to overcome the current limitations of datasets used for training.