This paper presents a method for learning to navigate in complex environments using deep reinforcement learning (DRL). The approach involves formulating navigation as a reinforcement learning problem and incorporating auxiliary tasks that leverage multimodal sensory inputs to improve data efficiency and task performance. The key idea is to jointly learn the goal-driven reinforcement learning problem with auxiliary tasks such as depth prediction and loop closure classification. This allows the agent to navigate from raw sensory input in complicated 3D mazes, achieving human-level performance even under conditions where the goal location changes frequently. The method uses a stacked LSTM architecture to address memory requirements and incorporates multiple objectives, including maximizing cumulative reward, minimizing an auxiliary loss for depth map inference, and detecting loop closures. The agent is trained using the Asynchronous Advantage Actor-Critic (A3C) algorithm, which learns both a policy and a value function. The agent's performance is evaluated on five 3D maze environments, demonstrating accelerated learning and increased performance. The agent's ability to localize and its network activity dynamics are analyzed, showing that it implicitly learns key navigation abilities. The paper also explores different variants of depth prediction, including regression and classification tasks, and finds that classification tasks yield better results. The method is validated in challenging maze domains with random start and goal locations. The results show that the agent can navigate efficiently, even in complex environments, and that its performance is correlated with its ability to localize. The agent's ability to detect loop closures is also demonstrated, with high accuracy in detecting true positives and minimizing false positives. The paper also compares the performance of different agent architectures, including feedforward models, recurrent models, and stacked LSTM models with additional inputs. The results show that the stacked LSTM model with depth prediction and loop closure prediction performs best. The paper also discusses the importance of auxiliary tasks in improving data efficiency and the potential of using external memory to enhance navigational abilities. The work highlights the effectiveness of using auxiliary tasks in deep reinforcement learning for navigation tasks.This paper presents a method for learning to navigate in complex environments using deep reinforcement learning (DRL). The approach involves formulating navigation as a reinforcement learning problem and incorporating auxiliary tasks that leverage multimodal sensory inputs to improve data efficiency and task performance. The key idea is to jointly learn the goal-driven reinforcement learning problem with auxiliary tasks such as depth prediction and loop closure classification. This allows the agent to navigate from raw sensory input in complicated 3D mazes, achieving human-level performance even under conditions where the goal location changes frequently. The method uses a stacked LSTM architecture to address memory requirements and incorporates multiple objectives, including maximizing cumulative reward, minimizing an auxiliary loss for depth map inference, and detecting loop closures. The agent is trained using the Asynchronous Advantage Actor-Critic (A3C) algorithm, which learns both a policy and a value function. The agent's performance is evaluated on five 3D maze environments, demonstrating accelerated learning and increased performance. The agent's ability to localize and its network activity dynamics are analyzed, showing that it implicitly learns key navigation abilities. The paper also explores different variants of depth prediction, including regression and classification tasks, and finds that classification tasks yield better results. The method is validated in challenging maze domains with random start and goal locations. The results show that the agent can navigate efficiently, even in complex environments, and that its performance is correlated with its ability to localize. The agent's ability to detect loop closures is also demonstrated, with high accuracy in detecting true positives and minimizing false positives. The paper also compares the performance of different agent architectures, including feedforward models, recurrent models, and stacked LSTM models with additional inputs. The results show that the stacked LSTM model with depth prediction and loop closure prediction performs best. The paper also discusses the importance of auxiliary tasks in improving data efficiency and the potential of using external memory to enhance navigational abilities. The work highlights the effectiveness of using auxiliary tasks in deep reinforcement learning for navigation tasks.