چكيده به لاتين
With science and technology development, the need for robots and their improvement in various fields is felt more than before. One of the most commonly used robots is wheeled mobile robots, which are planetary exploration robots, agricultural machinery, object-moving robots, Relief robots, and other kinds of that. The choice of controller type is based on the type of system used in terms of linearity and nonlinearity, the desired goal, environment conditions, and constraints related to the issue. Improving these controllers to improve accuracy and solving speed requires more attention, and there are essential points both in terms of software and coding and in terms of hardware that can be applied and make better performance of controllers. This thesis examines the weaknesses, proposes solving them, applying the solutions, and compares the results in simulation and practical testing. For this purpose, a four-wheeled mobile robot is designed for simulation and implementation. The Linear Quadratic Regulator (LQR) controller and the Nonlinear Model Predictive Control (NMPC) are used to control it. Recent research has shown that combining modern and classic controllers with machine learning can lead to better performance of robotic systems. So, for increasing accuracy of path tracking, this controllers become intelligent with Multi-Layer Perceptrons Neural Networks (MLP-NN). The use of controllers with a cost function, despite creating optimization or resilience or predictability in the design goal; adjusting its gains, which are used to weight the system state vector as well as the system input vector, is a too difficult and time-consuming task that is usually chosen by trial and error method for each specific application. In addition, with the slightest change in the gains, problem-solving and system optimization are severely affected, and therefore it is necessary to select these gains intelligently to improve controller performance. In order to increase the performance accuracy of these controllers, the combination of controllers based on cost function with Neural Network (NN) has been done; by presenting and using a new algorithm in network training, smart optimal gains are extracted in each step and used to minimize the cost function. Also, in order to reduce the existing time delays, especially in the implementation of the nonlinear controller on the robot, by training another neural network to optimally extract the predictive horizon gain, reduce the calculations and increase the solution speed. By applying this network, not only the speed of solving but also the accuracy of tracking has increased. Furthermore, in the hardware section, time delays have been reduced by examining and using tools with higher speeds using Pixy2 camera and U2D2 for data transportation.
