designing a sliding mode control system for a hovercraft and improving it with deep reinforcement learning

The control set in most moving vehicles in water faces an interferential and nonlinear system. specifically, in a hovercraft, due to the assignment of underactuated and insufficient actuators, the control effect is highly interferential in the channels. in this vehicle, the dynamics change significantly in each maneuver (speed). with rudder deflection in sway motion, the surge channel is affected, and similarly, with a change in surge velocity, the sway channel behavior is completely transformed. in this study, by identifying the desired behavior, initially, a sliding mode controller with the requirement of minimal chattering in commands for the surge and sway velocity channels is designed. the sliding mode controller is a robust controller whose stability can be proven. since this controller is designed for a specific system characteristic of the hovercraft and with conservative variables, it does not necessarily exhibit suitable behavior with small tracking error for all maneuvers and uncertainties. inevitably, using reinforcement learning and the ppo method, the initial controller is adjusted for most possible states with the constraint of reducing chattering and increasing tracking accuracy. uncertainties in system characteristics and motion maneuvers are modeled in the simulation program and used in the learning. learning calculations are performed offline. the result is applied as a trained actor-critic neural network to the initial sliding mode controller. the research results show that by tuning the controller with machine learning, precise commands are executed without large oscillations being introduced into the control. additionally, the average cumulative reward increases by at least 40%.