Foveation Control of a Robotic Eye Using Deep Reinforcement Learning

Abstract

Deficit of the extraocular muscle is known as a key cause of ocular motility disorders that affect eye movement and complicate daily activities of millions of people in the US. A physical model mimicking the biomechanics of the oculomotor plant can improve the understanding of functionality and control of extraocular muscles and provide a tool for researchers to gain insights into binocular misalignment. This paper will present, for the first time, the design and development of a robotic eye system driven by antagonistic super coiled polymer (SCP) based artificial muscles and the motion control design by leveraging machine learning techniques. The dynamic model of the robotic eye will be presented. Deep reinforcement learning is used for control design of the robotic eye system, demonstrated by simulation of one-dimensional foveation control.

Two degree-of-freedom robotic eye: design, modeling, and learning-based control in foveation and smooth pursuit

“…The dynamic model of the 2-DOF robotic eye is derived from its geometric model and the dynamics of the SCP muscles. Following [14,15], the SCP muscle dynamics depends on independent state variables including the change and the changing rate of the length of the muscles as well as the change of temperatures of the muscles. While SCP muscles also demonstrate hysteresis behaviors, the corresponding performance loss has been studied and can be eliminated by inverse compensation [24].…”

“…Table 1 shows the parameters used in the case studies. Here, the dynamic parameters of the SCP actuator follow [14,15]. While there may be differences in the parameters of an SCP muscle each time it is fabricated, the average values of the parameters are adopted.…”

“…However, the rigidity and motor-based control make such designs differ from the actual biomechanics of the human eye [5], which rotates by actuations of the extraocular muscles. In recent advancement, artificial muscles, such as pneumatic actuators [11][12][13], and super coiled polymers (SCPs) [14,15], have been adopted to provide more flexibility and closer mimicry of the human eye.…”

“…This paper focuses on the design, modeling and control of a robotic eye using SCP artificial muscles. SCP is a newly developed artificial muscle that has a high power-to-weight ratio, is cost effective, and provides a large contraction strain [14,15], thus making it a strong fit for the development of a robotic eye model. This paper presents the design and dynamic modeling of a two degree-of-freedom (2-DOF) robotic eye system, and further investigates the foveation and smooth pursuit motion.…”