The need for autonomous vehicles that do not operate on highways and can move off-road is increasing. While there have been significant advancements in autonomous passenger vehicles in the last couple of years, autonomous off-road vehicles have not received as much of this attention. The demand to deliver emergency supplies to areas unreachable by typical highways advocates the need for an off-road self-driven vehicle. This dissertation discusses transforming an All-Terrain Vehicle into a self-driven off-road vehicle that can follow a path in a forest environment. The approach is to install actuators, sensors, microcontrollers, and a graphical processing unit on top of the ATV without changing the initial ATV architecture. This research was composed of two phases. The first phase was to augment the ATV with actuators and sensors to have a reliable base where digital signals control the ATV. This phase focused on controlling the handlebar of the ATV in a forward direction, having a feedback speed control system, and implementing a braking control system. The second phase involved developing an Artificial Intelligence model with an image processing system to compute the proper steering angle and control the ATV. This phase consisted of an OpenCV module to read a camera feed and pass it to a semantic segmentation deep learning model called SegNet, which helps identify the paved path. The software received the semantic segmentation output data identified by the ATV's current position on the paved path and the expected trajectory of the ATV. A trained Machine Learning Linear Regression model used this data to predict the handlebar angle, pass the angle to a CAN bus to steer the ATV, and keep the ATV on the paved path. The dissertation includes the data gathering process, the ML model training, and the LR model choice. The research also included a testing method to check the ATV self-driving performance when following a paved path in the forest referencing an experienced human driver. The testing model adopts an Inverse Reinforcement Learning model that learns the best driving policy from an experienced driver in the environment of the paved path in the forest and outputs a reward function. The camera feeds the environment into a semantic segmentation model. The output reward function delivered a rating equation that can be applied to other driving techniques to measure the performance.