Autonomous Mapless Navigation via Maximum Entropy Learning

Autonomous mapless navigation is a significant problem for autonomous robots when the entire environment is unknown. In recent years, an increasing number of methods based on deep reinforcement learning have been used for autonomous navigation. However, some of these methods are closely tied to the training environment, making it difficult for them to be generalized to mapless environments. Moreover, researchers are often confused by the problem of easily getting trapped in the local optimum zone. To address these issues, this paper proposes a method based on deep reinforcement learning that 1) takes the neighboring environment as input to the agent, ensuring that the policy learned by the agent can be generalized to mapless environments and 2) replaces maximum value learning with maximum entropy learning to improve the agent's exploration ability. The final results show that our method can autonomously navigate in mapless scenarios.