Modeling and Learning-Based Control for Super-Coiled Polymer-Driven Robotic Eye

A newly-developed type of artificial muscles – Super-Coiled Polymer (SCP), comparatively offers many advantages in terms of cost, size, flexibility, fabrication, and power-to-weight ratio, potentially making SCPs a great fit for deployment in bioinspired robots. Development of bioinspired robots incorporating artificial muscles, increasingly necessitates derivation of precise dynamic models for motion prediction and controller design. Nevertheless, the process of modeling the system dynamics of such sophisticatedly evolving robots becomes difficult due to their continuum dynamics and high dimensionality. To address the problems of high nonlinearity and intrinsically infinite system dimension, contemporary artificial intelligence techniques, specifically reinforcement learning algorithms, are employed to design learning-based controllers. This necessity of developing intelligent control serves as the motivation to not only mimic the biomechanisms, but also mimic the cognitive abilities of these biological life forms, which is where learning-based controllers will play a major role in such bioinspired robotic systems. Our research aims at designing and developing a motivating SCP-driven bioinspired robotic eye, which aims to aid ophthalmologists, ocularists, and biomedical researchers in understanding better, the movement of the extraocular muscles of the human eye. Consequently, this could give comprehensive insights in studying the neuro-biomechanics of oculomotor disorders such as misalignment of the eyes in patients with strabismus, thus enabling diagnostic researchers to profoundly investigate the fundamental cause and effect of such disorders. Concurrently, we focus on modeling its system dynamics, and developing a robust learning-based controller to achieve various objectives such as orientation and perceptive control. Our proposed learning-based control design employs the use of deep-deterministic policy gradient (DDPG) reinforcement learning algorithm to train the agent with a linear quadratic regulator (LQR) based multi-objective reward. The effectiveness of the proposed control method was verified through simulations by performing tests of ocular foveation and smooth pursuit.