Target Tracking for Muscle-Skeleton Robotic Arm Using Large Language Model-Enhanced Reinforcement Learning

Human biological structures and neural mechanisms have provided many great inspirations for the design and improvement of human-like robots. McKibben muscle, as a new type of flexible actuator, can be used to manufacture humanoid robots. In this article, we use a neural network combining a Kinect v2 camera to predict air pressure to realize the high accuracy control of a muscle-skeleton robotic arm. The robotic arm, which is made of multifilament muscles, can make movements similar to human arms. And its performance is verified in the simulation environment using the reinforcement learning method. In the above process, we use the large language model (LLM) to enhance reinforcement learning and neural network training. The experimental results show that the LLM can help us train a suitable controller for the robotic arm based on experience. And the robotic arm can complete the spatial target-tracking tasks with an average error of 2.14 cm.