Reinforcement Learning Control of Jellyfish-Inspired Robot

Abstract

In this paper, a reinforcement learning (RL) based control method is presented, which allows a biomimetic jellyfish robot driven by a DC motor to track several trajectories. The novel jellyfish robot is fabricated to mimic the locomotive behavior of jellyfish and is capable of executing vertical propulsion and displacement control. Eight bells with curved shapes are actuated with a built-in DC motor. Rapid rotary movement is generated by the DC motor, to provide the thrust force to the jellyfish robot. Through a Scotch-Yoke mechanism, the output of rotary motion of the DC motor is converted to linear vertical reciprocating motion. Compared with traditional flexible mechanical jellyfish, the mechanical jellyfish driven by a DC motor has a higher speed with better controllability. Through independent and coordinated control of the driving voltage and swinging frequency, the jellyfish robot replicates vertical movement at various speeds and stability, similar to the real jellyfish. Due to nonlinear fluid dynamics and electromechanical coupling, a RL-based control plus a model-free PID control (RL-PID) is used to offer powerful algorithms for finding optimal controllers. This enables the jellyfish robot to follow trajectories. Experimental results have verified the effectiveness of the RL-PID control.