Navigating in unknown environments represents a significant challenge for robotics and autonomous systems. This problem fits into the more general context of path planning and it is even more complex when the environment and the controlling actions must be represented by continuous variables. Deep Reinforcement Learning is a possible solution even if it becomes challenging for lack of generalization in unknown scenarios. In this paper, using a Soft Actor-Critic approach, two possible methodologies for navigating in unknown environments are analyzed. To test the proposed models in situations not observed during the training phase, we also introduce a synthetic dataset through which we evaluate the performance.