Data-Driven Twisted String Actuation for Lightweight and Compliant Anthropomorphic Dexterous Hands

Abstract

Anthropomorphic dexterous hands are crucial for robotic interaction in unstructured environments, yet their performance is often constrained by traditional actuation systems, which suffer from excessive weight, complexity, and limited compliance. Twisted String Actuators (TSAs) offer a promising alternative due to their high transmission ratio, lightweight design, and inherent compliance. However, their strong nonlinearity under variable loads poses significant challenges for high-precision control. This study presents an integrated approach combining data-driven modeling and biomimetic mechanism innovation to overcome these limitations. First, a data-driven modeling approach based on a dual hidden-layer Back Propagation Neural Network (BPNN) is proposed to predict TSA displacement under variable loads (0.1-4.2 kg) with high accuracy. Second, a lightweight, underactuated five-finger dexterous hand is developed, featuring a biomimetic three-phalanx structure and a tendon-spring transmission mechanism, achieving an ultra-lightweight design. Finally, a comprehensive experimental platform validates the system's performance, demonstrating precise bending angle prediction (via integrated BPNN-kinematic modeling), versatile gesture replication, and robust grasping capabilities (with a maximum fingertip force of 7.4 N). This work not only advances TSA modeling for variable-load applications but also provides a new paradigm for designing high-performance, lightweight dexterous hands in robotics.