Implementation of Neural Networks for Plant Model Optimization in Predictive Controllers

• Grant number: 24/09740-1
• Support Opportunities: Scholarships in Brazil - Scientific Initiation
• Start date: November 01, 2024
• End date: October 31, 2025
• Field of knowledge: Engineering - Mechanical Engineering
• Principal Investigator: Marcelo Becker
• Grantee: João Henrique Alessio
• Host Institution: Escola de Engenharia de São Carlos (EESC). Universidade de São Paulo (USP). São Carlos , SP, Brazil

Abstract

Currently, the control of high-precision robotic systems has been the subject of numerous studies. In this context, Model Predictive Controllers have gained prominence in the automation of autonomous vehicles. The optimization of the control process is commonly the focus of scientific approaches. However, to achieve optimal predictive control, it is equally necessary to have a dynamic model that accurately represents the system's reality.The modeling of the plant in robotic systems is typically considered a geometric problem, involving extensive and complex calculations to represent the robot's dynamics and kinematics. However, a purely geometric perspective of the project can be insufficient for many problems where there is limited knowledge about the system's functioning. In nonlinear processes, for example, it becomes impossible to perform the simplifications commonly used in the conventional method. Therefore, it is essential to go beyond geometric approaches to achieve greater realism.In this context, approximation methods using neural networks have gained traction in the field of System Identification. These methods enable supervised or unsupervised learning of the relationships between the system's inputs and outputs based on the robot's previous experiences. By using the model obtained in a predictive control framework, it is possible to achieve more accurate trajectory predictions for the vehicle.The modeling of the plant in robotic systems is typically considered a geometric problem, involving extensive and complex calculations to represent the robot's dynamics and kinematics. However, a purely geometric perspective of the project can be insufficient for many problems where there is limited knowledge about the system's functioning. In nonlinear processes, for example, it becomes impossible to perform the simplifications commonly used in the conventional method. Therefore, it is essential to go beyond geometric approaches to achieve greater realism.In this context, approximation methods using neural networks have gained traction in the field of \textit{System Identification}. These methods enable supervised or unsupervised learning of the relationships between the system's inputs and outputs based on the robot's previous experiences. By using the model obtained in a predictive control framework, it is possible to achieve more accurate trajectory predictions for the vehicle.