Navigating Partially Unknown Environments: A Weakly Supervised Learning Approach to Path Planning

Abstract

In fire rescue missions, the critical research concern revolves around enabling autonomous path planning for mobile robots to quickly and safely navigate to target points. This paper focuses on sampling-based path planning methods under weak supervision. In order to enhance path quality and computational speed, we employ deep learning to perform non-uniform sampling on sampling-based methods, focusing on regions where optimal paths are more likely to exist. Specifically, the generation of non-uniform sampling regions is regarded as a semantic segmentation problem. In this context, diverse map information is utilized to predict non-uniform sampling regions. Inspired by attention mechanisms in deep learning, we propose an attention-guided model for non-uniform sampling path planning. The learning-driven path planning process comprises offline dataset generation, model training, and online model prediction. However, the offline dataset generation is often time-consuming and resource-intensive. To address this challenge, we propose a weakly supervised strategy, which necessitates the generation of only one single path as ground truth per scenario in semantic segmentation training. Furthermore, considering the potential existence of unknown obstacles along the reference path in real-world settings, we leverage deep reinforcement learning to ensure collision-free path tracking in unknown environments. Finally, extensive experimental simulations are conducted to verify the performance of the attention-guided model and collision-free tracking, and demonstrate the superiority of our proposed weakly supervised strategy.