Today, with the improvement of artificial intelligence and data-driven methods, in addition to conventional motion planning methods of robot, which are based on computational techniques, data-driven and machine learning-based motion planning methods are also used. In various industries, solving the challenge of pathfinding and motion planning for two-wheeled mobile robots, in order to move along a safe path without colliding with environmental obstacles, and reaches the target point by traversing the chosen path, has special importance. In this thesis, a trained model is used that utilizes the collected dataset to predict the desired path. For the mentioned model, a Convolutional Neural Network (CNN) model is proposed that has the capability to learn the complex mapping of sensor data and the target point information vector (the endpoint of the path) to the desired kinematic quantities. The convolutional network is employed in a hybrid control algorithm, which includes classical control methods and intelligent control. After generating the desired kinematic quantities using the intelligent feedforward component (CNN), a classical discrete-time feedback controller is used to converge the computational steps to the desired quantities. The problem-solving method with the proposed approach in this research, consists of seven main steps. In the first step, a set of algorithms is used to collect data for training the convolutional network model based on supervised learning. First, by generating a map of the environment and using the PRM path planning algorithm, the desired path is designed without considering the time factor. Then, using the kinematic equations of the robot and the pure pursuit controller, which is a geometric path tracking controller, the movement of a two-wheeled robot to follow the designed path is simulated. During path tracking by the robot, the sensor data and the target information vector are stored as input (features) of the dataset, while the linear velocity and the angular velocity are stored as output (labels) of the dataset. In the second step, by experimental selecting the structure of the CNN, the designed model is trained using the dataset collected in the first step. In the third step, after training the network and finding the optimal parameters of model, the model's performance is eva‎luated using the test dataset. In the fourth and fifth steps, the post-processing procedure is applied to the network's predictions, which includes converting the velocity kinematic quantities to position kinematic quantities and interpolating them, in order to compute the set of desired kinematic quantities for the input of the sixth step. In the sixth step, the intelligent feedforward component is implemented, which uses the output from the previous step and inverse dynamics to generate the required forces to achieve the desired kinematic quantities. In the final step of this algorithm, the state-space description of the system is linearized using the feedback linearization method, and then the position error of the robot is reduced using state feedback control. After convergence in the inner feedback loop, the algorithm returns to the starting point for the next time step using an outer loop, so that ultimately, the path predicted by the network is followed by the robot. The results of this thesis show that the hybrid control algorithm, which includes the trained network model, inverse dynamics, and feedback control, can guide the robot to reach the endpoint through obstacles in both the training environment and new environments. In general, in the present thesis, the dataset is collected through simulation to train the convolutional network model, and after training the network model, the hybrid algorithm consisting of the CNN, inverse dynamics, and feedback control is employed with the aim of pathfinding and enabling the robot to reach the endpoint in an environment with obstacles.

حميدرضا ميردامادي

سعيد ضيائي راد , مرضيه مجدراصيل