This paper explores the impact of reward mechanisms on the performance of an agent navigating an unknown environment. The study aims to determine which reward formulation best enables a mobile robot to develop goal-reaching and obstacle-avoidance behaviors. We employ Deep Q-Network (DQN) and Deep Deterministic Policy Gradient (DDPG) algorithms to assess how an agent’s available actions influence its movement and performance metrics. The agent, relying solely on laser distance readings, is tested in a simulated environment for goal-driven navigation. We evaluate DQN and DDPG under both sparse and dense reward functions. Experimental results indicate that DQN successfully learns goal-reaching behavior with both reward types and outperforms DDPG in terms of average reward, success rate, and collision avoidance. In addition, DQN demonstrates greater robustness to variations in reward function design. Our findings highlight that when using sparse laser scans as the state representation, the DQN agent consistently outperforms the DDPG agent across all reward configurations. Notably, even with a sparse reward, the DQN agent effectively reaches its goal using the same policy architecture.