Using Inverse Learning for Controlling Bionic Soft Robot Fish with SMA Actuators

In our previous research, we developed an untethered bionic soft robot fish for swimming motion. The body of the fish was molded by using soft silicone rubber, and we introduced shape memory alloy (SMA) wires for its actuators. The lightness and flexibility allow the robotics fish to generate biomimetic swimming motions. However, due to the hand-production process, asymmetrical structure of the fish's body cannot be avoided, which reduces the controllability of the bionic fish, especially when generating complex movement. Due to the complexities of creating an exact mathematical model of the swimming motions of the fish having the soft body, building a simulator for offline learning is prohibitively difficult. In this study, we introduce offline inverse learning with a feedforward neural network to generate control parameters for realizing desired swimming motion.<br/>The offline inverse learning refers to the training of a feedforward neural network to establish relationships between control parameters (tail swing frequencies, duty ratio of flipping duty cycle) and physical motion characteristics (speed, acceleration, direction of motion, body inclination angle) of robotic motion. The training data for inverse learning are obtained by measuring and observing the motions of the robot fish in relation to various control parameter combinations. Here, we randomly generate various control parameters (range of swing tale frequency: 0.25-5Hz, range of duty ratio for the input current to SMA: 0%-100%), and measure to examine the movement features for each set of parameters. The moving features are obtained from the fish's Inertial Measurement Unit (IMU) module, and also manually extracted from video of the fish's movements. After acquiring the data, they are inverted in that the movement features become the input and the control parameters become the desired output for a feed forward neural network. Once the neural network is trained, it can be utilized for controlling the desired maneuvers of the robot. Here, a motion is divided into a sequence of sub-motions and the neural network can be utilized to generate the control parameters of each sub-motion in a sequential manner. The efficiency of the proposed control mechanism will be evaluated in the physical environment through some pre-designed maneuvers.