Humanoid Muscle-Skeleton Robot Arm Design and Control Based on Reinforcement Learning

Abstract

Muscle-skeleton robots share similar appearances and functions with humans, making these robots more adaptive in human interaction scenarios. In this paper, a new muscle-skeleton robot arm driven by artificial muscles is proposed. First, we design a new multifilament McKibben muscle and measure its properties. Then a humanoid robot arm referred to the anatomy of the human arm is presented, while the configuration of muscle is adjusted to reduce the complexity of manufacturing and controlling. Muscle-skeleton robot arms with different muscle configurations are controlled using the reinforcement learning method in the simulation environment, and different arm models' movement ranges are obtained to find the best muscle configuration. The experimental results show that the model with the best muscle configuration achieves 79.8% of the whole movement range.