In this project we use several algorithms to tackle the problem of controlling an autonomous vehicle in the tasks of lane keeping and crash avoidance. The main simulation environment is ROS where we implemented a vehicle model equipped with a realistic laser sensor and a camera. Two separate simulation scenarios are designed for each task, one scenario is for training the algorithms and the other is to test the performance of the trained models. Three main learning algorithms are used. A discrete reinforcement learning algorithm called Q-learning, A continuous reinforcement learning algorithm called DDPG. The previous algorithms used laser sensor readings as input. The final algorithm, which is a supervised deep learning algorithm that uses an architecture of a Convolutional Neural Network, relied on images fed from the camera as input. A comparative study of the results is done based on specific evaluation criteria.

Read the report (graduation-design-project.pdf) for detailed description of the work.