Data-Driven Performance-Prescribed Reinforcement Learning Control of an Unmanned Surface Vehicle

Area of Science:

• Robotics and Control Systems
• Marine Engineering
• Artificial Intelligence

Background:

• Traditional model-based control methods are often infeasible for unmanned surface vehicles (USVs) operating in complex, unpredictable marine environments due to difficulties in accurate system modeling.
• Existing control strategies may struggle to simultaneously achieve optimal control performance and guarantee specific tracking accuracy, especially under unknown system dynamics.

Purpose of the Study:

• To develop an innovative, model-free control solution for USVs that relies solely on input-output signals.
• To create a data-driven control scheme that integrates reinforcement learning with prescribed performance to achieve both control optimality and predefined tracking accuracy.
• To enable robust USV navigation and control in challenging marine conditions without prior system identification.

Main Methods:

• A data-driven performance-prescribed reinforcement learning control (DPRLC) scheme was developed.
• State transformation with prescribed performance was employed to convert constrained tracking errors into unconstrained stabilization problems, even with unknown USV dynamics.
• A neural network-based actor-critic reinforcement learning framework was utilized to optimize controller synthesis directly from the Bellman error formulation.

Main Results:

• The proposed DPRLC scheme successfully converts tracking error constraints into a stabilization problem for unknown dynamics.
• The reinforcement learning approach optimizes controller synthesis, leading to a data-driven optimal controller.
• Theoretical analysis confirmed that the DPRLC scheme guarantees prescribed tracking accuracy while optimizing control cost.

Conclusions:

• The developed DPRLC scheme offers a highly effective and superior model-free solution for USV control in complex marine environments.
• The approach demonstrates remarkable effectiveness in achieving both optimal control and prescribed tracking accuracy, validated through simulations and virtual-reality experiments.
• This self-learning strategy overcomes the limitations of model-based approaches, paving the way for more autonomous and capable USV operations.