The long-standing debate about pros and cons of combining learning and evolution seems endless, despite the huge amount of research works. Nonetheless, some authors recently employed Lamarckian learning in combination with evolutionary algorithms to synthesize successful controllers in different domains, in which both morphology and control of the robots can evolve. Notwithstanding the amount of efforts, the effects of combining of evolution and learning in autonomous robots still need to be deepened. In this work, a quantitative and qualitative analysis about whether and how learning provides benefits to evolution is presented, in which a robotic task involving the interaction of the agent with the environment is considered. Specifically, a foraging task is proposed, in which the robot has to discriminate between different food items providing diverse rewards. A pure evolutionary algorithm is compared with a new method, called EVoLE, which combines evolution and Lamarckian learning implemented through back-propagation. Results collected in this domain indicate that learning actually enhances evolution. In fact, the application of back-propagation provides a significant advantage in terms of performance. Moreover, the analysis of the displayed behaviors reveals how learning enables the development of qualitatively superior strategies than those discovered by evolution.