Learning Adaptive Mapless Navigation Skills Through Evidential Deep Learning

Traditional navigation methods tend to design and train navigation skills based on the consideration of multi-scenario applicability and perform well in general scenarios. However, this results in sub-optimal navigation performance in some context-specific scenarios, such as navigating through open or crowded scenarios. While some existing navigation skills can achieve satisfactory performance in context-specific scenarios, e.g., using PID in open scenarios and reinforcement learning-based training methods in crowded scenarios. Most methods currently do without consideration of the integration of navigation skills and contextual understanding. To address this gap, we propose an adaptive mapless navigation method to utilize the optimal capabilities of each skill corresponding to the context-specific scenario. Specifically, we package several existing navigation skills and train a scenario classifier to enable the agent to automatically select the most appropriate skill based on the current scenario. Furthermore, we incorporate evidential deep learning (EDL) to assess classification uncertainty and ensure the safe switching of navigation skills. Our method was tested in both simulated and real scenarios, and the results of the experiments confirmed the effectiveness and generalization of our method. Moreover, the proposed method exhibits notably enhanced, context-sensitive efficiency within a hybrid multi-scenario setting, in comparison to a single navigation skill.