Design, Modeling, and Visual Learning-Based Control of Soft Robotic Fish Driven by Super-Coiled Polymers

A rapidly growing field of aquatic bio-inspired soft robotics takes advantage of the underwater animals' bio-mechanisms, where its applications are foreseen in a vast domain such as underwater exploration, environmental monitoring, search and rescue, oil-spill detection, etc. Improved maneuverability and locomotion of such robots call for designs with higher level of biomimicry, reduced order of complex modeling due to continuum elastic dynamics, and challenging robust nonlinear controllers. This paper presents a novel design of a soft robotic fish actively actuated by a newly developed kind of artificial muscles-super-coiled polymers (SCP) and passively propelled by a caudal fin. Besides SCP exhibiting several advantages in terms of flexibility, cost and fabrication duration, this design benefits from the SCP's significantly quicker recovery due to water-based cooling. The soft robotic fish is approximated as a 3-link representation and mathematically modeled from its geometric and dynamic perspectives to constitute the combined system dynamics of the SCP actuators and hydrodynamics of the fish, thus realizing two-dimensional fish-swimming motion. The nonlinear dynamic model of the SCP driven soft robotic fish, ignoring uncertainties and unmodeled dynamics, necessitates the development of robust/intelligent control which serves as the motivation to not only mimic the bio-mechanisms, but also mimic the cognitive abilities of a real fish. Therefore, a learning-based control design is proposed to meet the yaw control objective and study its performance in path following via various swimming patterns. The proposed learning-based control design employs the use of deep-deterministic policy gradient (DDPG) reinforcement learning algorithm to train the agent. To overcome the limitations of sensing the soft robotic fish's states by designing complex embedded sensors, overhead image-based observations are generated and input to convolutional neural networks (CNNs) to deduce the curvature dynamics of the soft robot. A linear quadratic regulator (LQR) based multi-objective reward is proposed to reinforce the learning feedback of the agent during training. The DDPG-based control design is simulated and the corresponding results are presented.