Bioinspired Reinforcement Learning Control for a Biomimetic Artificial Muscle Pair

Artificial muscles are recently developed actuators extremely promising for compliant robotic systems. Their accurate closed-loop control is challenging due to their highly nonlinear behavior. In this work, we model an artificial muscle pair adopting a non-pulley configuration mimicking more realistically the behavior of biological muscles. Inspired by how the brain regulates dopamine-based learning from interaction with the environment, it is possible to design efficient reinforcement learning control algorithms. Therefore, we propose a reinforcement learning-based controller bioinspired by the parallels between the behavior of temporal difference errors and the activity of dopaminergic neurons. Simulated experiments conducted in a virtual scenario show that the control action can accurately tackle the nonlinear control problem. The proposed solution could be extended to the dynamic control of more realistic and complex anthropomorphic limb systems due to its inherent adaptability and control effectiveness regardless of the complexity of the environment.