Neural-Inspired Model-Based Reinforcement Learning Method for Motion Control of Musculoskeletal Robot

Controlling musculoskeletal robots to reach target positions with high precision is an important research direction for this type of redundant system. The neural-inspired methods are beneficial to improving the performance of controllers by drawing lessons from the mechanisms of brain motion regulation. Inspired by mechanisms of the spatial cognition of the hippocampus and the sensorimotor modulation of the cerebellum, a model-based reinforcement learning method that integrates these two brain regions is proposed. This method adopts the model predictive control framework. Firstly, in terms of model learning, the hippocampus network and cerebellum network are designed based on the structures and functions of these brain regions. The hippocampus model uses a probabilistic neural network to segmentally learn the transition model of the musculoskeletal robot, and predict the system state. The cerebellum model extracts the feature vector of historical transitions to improve the learning ability of the hippocampus model. Secondly, in terms of model utilization, the cross entropy method is used to solve the muscle control signals. The proposed method is verified on a musculoskeletal robot that mimics a humanoid arm. Experimental results show that the neural-inspired method can control the end-effector of the robot to reach random targets with fast learning efficiency and high motion precision.