基于数据增强的神经路径规划器

Neural Path Planner Using Data Augmentation

摘要: 传统的路径规划算法主要是基于搜索或采样的方法，如D*、A*、RRT、RRT*。但随真实应用场景日趋复杂，尤其是在3D复杂场景下，传统算法的计算量大幅增加，难以实现实时计算。近年来，鉴于深度神经网络在的快速推理能力以及非线性建模方面的优势，基于神经网络的路径规划算法成为目前路径规划研究的热点。本文针对神经路径规划器训练数据不易收集，数据不足的问题，提出了两种数据增强方法，分别在环境编码器和神经规划器训练阶段应用，以此达到丰富训练样本的目的，今模型学习到更多的知识。实验表明，本文所提方法较于之前的方法，在路径规划成功率以及时间两个评估指标上都取得优势，证明了本方法的优越性能，满足路径规划任务的需求。

Abstract: Traditional path planning algorithms are mainly based on search or sampling methods, such as D*, A*, RRT, and RRT*. However, with the increasing complexity of real application scenarios, especially in 3D complex scenarios, the computation of traditional algorithms increases dramatically, making it difficult to realize real-time computation. In recent years, given the advantages of deep neural networks in terms of their fast reasoning ability and nonlinear modeling, neural network-based path planning algorithms have become the current hotspot in path planning research. This paper introduces two data augmentation methods to address the challenges of limited and difficult-to-collect training data for neural path planners. These methods are implemented during the training phase of the environment encoder and the neural planner, respectively, with the aim of enhancing the training samples and improving the model's knowledge acquisition. Experimental results demonstrate that the proposed method outperforms previous approaches in terms of path planning success rate and time, thereby confirming its superior performance and suitability for path planning tasks.