Given its complexity within the upper limb, the shoulder joint presents a major design challenge in the creation of humanoid robots, driving the need for bio-inspired mechanisms that achieve both dexterity and stability. This study deals with the bio-inspired design, mechanical modeling and neural network-enhanced control of a novel cable-driven shoulder mechanism. The design of the cable-driven biomimetic shoulder mechanism is informed by the identification of a core muscle subset, which is determined by establishing the muscular simplification strategy through the analysis of muscle functions in representative tasks. Then, the cable force distribution of the mechanism is formulated as a constrained quadratic programming problem to optimize both energy efficiency and load uniformity. Subsequently, a prototype platform is constructed and experimentally evaluated, achieving approximately 80% coverage of the human shoulder joint workspace, with a mean motion error of 3.12° and repeatability below 1°. To improve its control performance, a cooperative control strategy based on state-transition learning was developed. This approach utilizes deep neural networks to model the dynamic relationship between cable tension and joint motion, which led to a 48.9% improvement in trajectory-tracking accuracy and a significant enhancement in tension consistency. These results collectively validate the efficacy of the proposed shoulder mechanism and offer valuable insights for the design of cable-driven humanoid shoulders and other bioinspired robotic joints.
Gao et al. (Wed,) studied this question.