The field of 3D perception and semantic segmentation is rapidly advancing with a focus on improving accuracy, efficiency, and robustness. Researchers are exploring new architectures and techniques to address challenges such as domain heterogeneity, limited training data, and adverse weather conditions. Notably, the development of novel frameworks and models, such as mixture-of-experts and cross-modal knowledge distillation, is enabling more effective utilization of multi-modal data and improving performance in various applications, including 3D object detection, semantic segmentation, and scene understanding. Noteworthy papers in this area include Point-MoE, which proposes a Mixture-of-Experts architecture for cross-domain generalization in 3D semantic segmentation, and SR3D, which introduces a training-free framework for single-view 3D reconstruction and grasping of transparent and specular objects. Other notable works, such as CroDiNo-KD and BiXFormer, are also making significant contributions to the field by leveraging disentanglement representation, contrastive learning, and modality-agnostic matching to improve performance and robustness.