Machine Learning for Vehicle Behavioural Control

With the advancement of Artificial Intelligence, a revolution in the automotive field has taken place in last few decades. With the development of Advanced Driver Assistance Systems (ADAS), various sensors were available, which paved theway for fully- autonomous vehicles. The first autonomous car has appeared in 1980's, since then a lot of research has taken place in order to develop an autonomous vehicle, which can drive the passengers all by itself without any human-driver intervention. This project work is aimed at developing a simulation model of an Autonomous Vehicle. The so developed Autonomous Vehicle is able to navigate through a pre-defined path avoiding obstacles, walkers, and other vehicles autonomously without any human control. For attaining this driving autonomy, a Reinforcement Learning Algorithm was developed and trained using Deep Q- Learning technique and in-loop-simulations on CARLA-simulator. Upon training the agent for 20,000 episodes of simulation, it is overt from simulation results that the learning accuracy of the agent aggrandized, the average reward nabbed by the agent amplified, the agent's training loss subsided, which reinforce that the agent started to acquire more positive average-reward than the negative average-reward. At the 17,150
^th
, episode of training, the agent concomitantly set forth peak learning accuracy, diminutive training loss and best average-reward. This tacit validate that the agent is adept in maneuvering without colliding with other vehicles, walkers, and with no lane invasions, autonomously from its experience earned from episodic in-loop-simulations.