Researchers propose a new framework for open-vocabulary 3D object detection that improves both the discovery of novel objects and the robustness of model training. Their co-distillation strategy distills high-quality novel objects by applying Hungarian matching to geometric consistency, structural objectness, and semantic certainty scores. To enhance robust model training, they also propose a dual-guidance learning scheme that incorporates scene-awareness-guided uncertainty regularization for regression heads and LLM-guided hierarchical alignment for classification heads. This approach is said to achieve significant performance gains over state-of-the-art methods in experiments on SUN RGB-D and ScanNetV2 datasets. --- Why it matters: This matters because it addresses a limitation of existing open-vocabulary 3D object detection approaches, where inaccurate localization and mismatched classification can limit model performance. The proposed framework has the potential to improve the accuracy and robustness of these models in real-world applications such as robotics and autonomous vehicles. Source: https://arxiv.org/abs/2608.19973