Bioinspired Gain-Modulated Recurrent Neural Network for Controlling Musculoskeletal Robot

摘要

The motor cortex can arouse abundant transient responses to generate complex movements with the regulation of neuromodulators, while its architecture remains unchanged. This characteristic endows humans with flexible and robust abilities in adapting to dynamic environments, which is exactly the bottleneck in the control of complex robots. In this article, inspired by the mechanisms of the motor cortex in encoding information and modulating motor commands, a biologically plausible gain-modulated recurrent neural network is proposed to control a highly redundant, coupled, and nonlinear musculoskeletal robot. As the characteristics observed in the motor cortex, this network is able to learn gain patterns for arousing transient responses to complete the desired movements, while the connections of synapses keep unchanged, and the dynamic stability of the network is maintained. A novel learning rule that mimics the mechanism of neuromodulators in regulating the learning process of the brain is put forward to learn gain patterns effectively. Meanwhile, inspired by error-based movement correction mechanism in the cerebellum, gain patterns learned from demonstration samples are leveraged as prior knowledge to improve calculation efficiency of the network in controlling novel movements. Experiments were conducted on an upper extremity musculoskeletal model with 11 muscles and a general articulated robot to perform goal-directed tasks. The results indicate that the gain-modulated neural network can effectively control a complex robot to complete various movements with high accuracy, and the proposed algorithms make it possible to realize fast generalization and incremental learning ability.