Q-learning is a machine learning technique that learns what to do and how to map states to actions to maximize rewards. Q-learning has been applied to various tasks such as foraging, soccer and prey-pursuing robots. In this paper, a simple foraging task has been considered to study the influences of the policies reported in the open literatures. A mobile robot is used to search and retrieve pucks back to a home location. The goal of this study is to identify an efficient policy for q-learning which maximizes the number of pucks collected and minimizes the number of collisions in the environment. Policies namely greedy, epsilon-greedy, Boltzmann distribution and random search are used to study their performances in the foraging task and the results are presented.