Researchers have developed CoToGrasp, a new AI framework for synthesizing grasps that are conditioned on specific contact topologies. Unlike existing grasp planners, which focus on physical stability, CoToGrasp learns to generate diverse and stable grasps in an object-agnostic manner. The model uses a canonical workspace to project local object features into a gripper-centric domain, allowing it to generalize to unseen objects without additional training data. According to the authors, CoToGrasp outperforms existing taxonomy-guided planners on the DexGraspNet dataset and has been demonstrated on a physical robot platform. --- Why it matters: This matters because current grasp planning methods often rely on object-annotated datasets, which can be expensive to collect. CoToGrasp's ability to generalize to unseen objects without additional training data could make it more practical for use in real-world applications. Source: https://arxiv.org/abs/2608.19776