Articulated objects are prevalent in daily life and robotic ma-nipulation tasks. However, compared to rigid objects, posetracking for articulated objects remains an underexploredproblem due to their inherent kinematic constraints. To ad-dress these challenges, this work proposes a novel point-pair-based pose tracking framework, termed PPF-Tracker. Theproposed framework first performs quasi-canonicalization ofpoint clouds in the SE(3) Lie group space, and then modelsarticulated objects using Point Pair Features (PPF) to predictpose voting parameters by leveraging the invariance proper-ties of SE(3). Finally, semantic information of joint axes is in-corporated to impose unified kinematic constraints across allparts of the articulated object. PPF-Tracker is systematicallyevaluated on both synthetic datasets and real-world scenar-ios, demonstrating strong generalization across diverse andchallenging environments. Experimental results highlight theeffectiveness and robustness of PPF-Tracker in multi-framepose tracking of articulated objects. We believe this work canfoster advances in robotics, embodied intelligence, and aug-mented reality.

Outdoor LiDAR point cloud 3D instance segmentation is acrucial task in autonomous driving. However, it requires la-borious human efforts to annotate the point cloud for train-ing a segmentation model. To address this challenge, wepropose a YoCo framework, which generates 3D pseudo la-bels using minimal coarse click annotations in the bird'seye view plane. It is a significant challenge to producehigh-quality pseudo labels from sparse annotations. Our YoCo framework first leverages vision foundation modelscombined with geometric constraints from point clouds toenhance pseudo label generation. Second, a temporal andspatial-based label updating module is designed to generatereliable updated labels. It leverages predictions from ad-jacent frames and utilizes the inherent density variation ofpoint clouds (dense near, sparse far). Finally, to further im-prove label quality, an IoU-guided enhancement module isproposed, replacing pseudo labels with high-confidence andhigh-IoU predictions. Experiments on the Waymo datasetdemonstrate YoCo's effectiveness and generality, achiev-ing state-of-the-art performance among weakly supervisedmethods and surpassing fully supervised Cylinder3D. Addi-tionally, the YoCo is suitable for various networks, achiev-ing performance comparable to fully supervised methodswith minimal fine-tuning using only O.8% of the fully la-beled data, significantly reducing annotation costs.