Virtual Attention Points Bridging Human Movement Characteristics And Dexterous Robot Motion Generation

In this study, we introduce Virtual Attention Points (VAPs) as a novel technique for characterizing the essence of dexterous human movements through mathematical encoding. This method focuses on pivotal points to capture movement dynamics, resulting in the generation of versatile and human-like motions for robotic systems. The proposed method inspired by the idea of human movement primitives (MPs) generates an interpretable low-dimensional representation for a given complex movement based on a new encoding basis function. Our approach achieves a remarkable 97% improvement in encoding accuracy for dexterous demonstrations with agile maneuvers and sharp turns, surpassing existing MP-based methods, enhancing the precision of fine manipulation and elevating the fidelity of encoded actions to the human movement. The precise replication of crucial poses and underlying behaviors highlights the efficacy of our approach in faithfully capturing the intricacies of expert human demonstrations. Our approach also generates a meaningful and interpretable repre- sentation of each demonstration, which encapsulates the skills-related features for performance assessment purposes. We propose a novel trajectory cloning algorithm that minimally warps various movement demonstrations such that starting and end points of motions will be ma- nipulated to desired locations. Our work holds transformative potential, especially in enhancing surgical training and autonomous surgical systems, where precision and human-like dexterity are paramount. As surgical operations necessitate dexterous trajectories to execute specific functional tasks like suturing, we implement the proposed method to assess its performance in surgical skills and autonomous surgery tasks.