Modeling and control of an artificial muscle pair through a bioinspired Reinforcement Learning algorithm

Artificial muscles are recently developed biomimetic actuators. They are extremely promising for providing naturally compliant robotic and mechatronic systems. However, accurate closed-loop control of these actuators is challenging due to their highly nonlinear behavior. In this work, a novel modeling and a control approach of an artificial muscle pair is described. The developed model is able to more accurately mimic the behavior of the human elbow musculature. Afterwards, a bioinspired reinforcement learning agent is trained. This type of learning is based on the parallels between the behavior of temporal difference errors and the activity of neurons that produce dopamine. The learned agent robustly tackles the nonlinear positioning control problem. This approach could be extended for the dynamic control of more realistic anthropomorphic limb systems due to its inherent adaptability and effectiveness regardless of environmental complexity.