The autonomous and intelligent robotics is related to different areas of knowledge, it is a multidisciplinary approach. The robots must be able, among other things, to predict and plan their actions. This is a learning process that requires long processing periods and several tests – which reinforces the importance of a virtual environment that is loyal to the intended real model and able to speed up the process of adjusting and learning model. The technique selected to provide intelligence to the virtual robot was Reinforcement Learning, where the agent (robot) receives rewards and punishments for actions taken. The aim of this technique is to find a policy that maximizes the rewards received by the agent. In this study, the Q-learning method was chosen. It is an off-policy method where the rewards’ maximization is independent of the policy used. The results were satisfactory, since the agent was able to learning in an unsupervised manner. The values of ε used in the ε-greedy function were essential to demonstrate the relevance of balance between exploration and exploitation in this kind of learning.