學習分層路徑規劃

路徑規劃對於許多任務至關重要，例如機器人導航和自動駕駛，在復雜的大型環境時，尋找正確的路徑往往會花費大量資源。因此，在本論文中，我們提出了一種階層式的路徑規劃方法以及一個通用的路徑規劃網路，我們將流程分為兩個階段: 整體規劃和局部規劃。在實行整體規劃時，我們調整地圖的大小並使用八個地圖來表示八個方向的障礙物分佈。在進行局部規劃時，我們關注在全局路徑中每個點所相對應的局部地圖。為了增加成功率，我們添加了重新規劃和重新選擇方向的機制。在實驗階段，我們評估訓練時間、執行時間和準確性。此外，我們展示了網絡的靈活性。

關鍵字

路徑規劃；分層式；有向圖；卷積神經網路；深度學習

並列摘要

Path planning is essential for many tasks, such as robot navigation and autonomous driving. When encountering a complex and large environment, finding the path costs a lot of resources. In this work, we introduce an efficient method with a general planning network. Instead of routing on the original map, we divide the process into two stages: global and detail routing. For the global routing, we resize the map and use eight maps to represent the obstacle distributions over eight directions. For detail routing, we concentrate on the local map corresponding to every point on the global path. To increase the success rate, we add re-routing and re-selection mechanisms. We evaluate our method by the training time, the execution time, and the accuracy. Furthermore, we show the flexibility of our network on action selection.