Bioinspired Flapping-Fin Unmanned Underwater Vehicles: Novel Deep Learning Inverse Control Methods to Optimize Efficiency, Propulsion, and Navigation Objectives on Live Constrained Autonomous Systems

Overview

Current approaches in autonomous robotics and control engineering lack practical efficiency optimization by prioritizing propulsion or relying on simulated data. Additionally, computationally expensive and slow deep learning approaches that could optimize robotic systems can’t fit constraints on live movement -by- movement control systems, which instead use preprogrammed movements or search a limited space of movements. I introduce a novel low -cost and flexible deep learning approach to control movement -by-movement objective optimization and integrate it fully on an PID -based unmanned underwater vehicle. Sensor data I collect trains forward neural networks that use the characteristic of a movement to predict the resultant propulsion and power consumption. Various models are tuned and benchmarked for computational/time performance on a Raspberry Pi and have a mean-squared error of 0.08%. An inverse search method invokes the forward model hundreds of times every half-second to search for an optimal movement by minimizing the created loss function, constantly repeating as the algorithm searches for a new movement based on the requested outcome by the controller. Communication is established between the Pi and PID systems, allowing physical testing to reveal a thrust propulsion improvement of 29% and reduction in power consumption by 66% with comp arable performance insights with other AV platforms through a dimensionless figure of merit. This expands the envelope of possible UUV missions including defense, climate research, and cleaning oil spills. On any system such as swarm robots, drones, or cars, our algorithm can adjust for prioritizing different objectives for low-power autonomous control, making efficient/optimal AV systems possible. Hawaii and Pacific