The field of dexterous grasping and hand manipulation is moving towards more generalizable and adaptable solutions. Researchers are exploring new methods to improve the robustness of vision-based grasping models, including the use of reinforcement learning and simulation-based data augmentation. Additionally, there is a growing interest in developing large-scale datasets that cover various grasp types and object categories, which is essential for training effective models. Another direction is the development of part-based methods for transferring hand-object interactions across different object categories, enabling more efficient learning and generalization. Furthermore, model-based reinforcement learning approaches are being investigated for learning multi-goal dexterous hand manipulation tasks. Notable papers in this area include:

• Dexonomy, which proposes an efficient pipeline for synthesizing contact-rich grasps for any grasp type, object, and articulated hand, and constructs a large-scale dataset with 10.7k objects and 9.5M grasps.
• PartHOI, which introduces a novel method for part-based hand-object interaction transfer using generalized cylinder representations, enabling robust geometric correspondence between object parts and cross-category transfer.