This research paper discusses the creation and application of a fully automated robot that follows a line by means of reinforcement learning techniques, more specifically the Q-learning algorithm. The system proposed comprises a vision-based perception module that uses OpenCV for its development and also includes the simulation environment of CoppeliaSim thus making the navigation process adaptive as well as robust. By interaction with its surroundings at all times, the agent gets trained on the control policy that is optimal in terms of lateral deviation minimization and hence provides stable trajectory tracking. The research comprises a detailed assessment of the Q-table dimensions and hyper-parameters such as the learning rate, discount factor, and exploration rate, systematically determining their impact on learning performance, convergence, and accuracy. According to the experimental findings, a configuration of a 7×7 Q-table strikes the best balance between precision and convergence speed that in turn results in smooth and even path tracking. The method, though quite effective under controlled conditions, does have its drawbacks in terms of state discretization, generalization, and real-time processing, thereby providing directions for the applications of deep reinforcement learning and adaptive perception models in the future.
Sekhar et al. (Mon,) studied this question.