This paper addresses the challenge of navigating in complex environments with dynamic elements, formulating the problem as a reinforcement learning task. The authors demonstrate that incorporating additional auxiliary tasks, such as depth prediction and loop closure classification, can significantly improve data efficiency and task performance. By jointly learning these tasks with the primary goal-driven reinforcement learning problem, the agents can navigate from raw sensory inputs in intricate 3D mazes, achieving human-level performance even when the goal location changes frequently. The paper provides detailed analysis of the agents' behavior, localization capabilities, and network activity dynamics, showing that the agents implicitly learn key navigation skills. The approach is evaluated using five 3D maze environments, featuring complex geometry, random start positions, dynamic goal locations, and long episodes. The results highlight the effectiveness of the proposed method in challenging environments, with agents demonstrating successful loop detection and efficient navigation strategies.This paper addresses the challenge of navigating in complex environments with dynamic elements, formulating the problem as a reinforcement learning task. The authors demonstrate that incorporating additional auxiliary tasks, such as depth prediction and loop closure classification, can significantly improve data efficiency and task performance. By jointly learning these tasks with the primary goal-driven reinforcement learning problem, the agents can navigate from raw sensory inputs in intricate 3D mazes, achieving human-level performance even when the goal location changes frequently. The paper provides detailed analysis of the agents' behavior, localization capabilities, and network activity dynamics, showing that the agents implicitly learn key navigation skills. The approach is evaluated using five 3D maze environments, featuring complex geometry, random start positions, dynamic goal locations, and long episodes. The results highlight the effectiveness of the proposed method in challenging environments, with agents demonstrating successful loop detection and efficient navigation strategies.