Path Planning using a One-shot-sampling Skeleton Map

TL;DR

This work tackles real-time path planning by learning a skeletonized representation of the free workspace to enable one-shot sampling. The authors introduce SkelUnet, a lightweight U-Net-based denoising autoencoder trained to reproduce skeletons of navigation maps, using Zhang-Suen outputs as targets. The SkelUnet-OSS framework builds a k-nearest-neighbor roadmap from the learned skeleton and answers multi-query path requests, enabling safer and faster navigation in 2D environments. Empirical results on 12,500 maps (with 250 unseen maps for testing) show improvements in safety-related metrics and processing time compared to traditional medial-axis approaches and A* baselines, with potential for extension to 3D environments. The work demonstrates that a learned skeleton can provide comprehensive workspace coverage with a smaller, more reliable graph for motion planning in UAV-like platforms.

Abstract

Path planning algorithms fundamentally aim to compute collision-free paths, with many works focusing on finding the optimal distance path. However, for several applications, a more suitable approach is to balance response time, path safety, and path length. In this context, a skeleton map is a useful tool in graph-based schemes, as it provides an intrinsic representation of the free workspace. However, standard skeletonization algorithms are computationally expensive, as they are primarly oriented towards image processing tasks. We propose an efficient path-planning methodology that finds safe paths within an acceptable processing time. This methodology leverages a Deep Denoising Autoencoder (DDAE) based on the U-Net architecture to compute a skeletonized version of the navigation map, which we refer to as SkelUnet. The SkelUnet network facilitates exploration of the entire workspace through one-shot sampling (OSS), as opposed to the iterative or probabilistic sampling used by previous algorithms. SkelUnet is trained and tested on a dataset consisting of 12,500 two-dimensional dungeon maps. The motion planning methodology is evaluated in a simulation environment with an Unmanned Aerial Vehicle (UAV) in 250 previously unseen maps and assessed using several navigation metrics to quantify the navigability of the computed paths. The results demonstrate that using SkelUnet to construct the roadmap offers significant advantages, such as connecting all regions of free workspace, providing safer paths, and reducing processing time.

Figures (8)

• Figure 1: Comparative path planning results across three indoor maps. Collision robot footprints are shown in green, while configuration within a collision radius are highlighted in orange. The planned path is indicated by a red line, and the executed trajectory is shown in blue. The first column is the trajectory using the skeleton obtained by SkelUnet the second row is the trajectory using the skeleton obtained by Zhang-Sue method and the third column is the trajectory computed with A* algorithm.
• Figure 2: SkelUnet-OSS planning scheme. The routine $\mathtt{SEEK\_CORNERS}$ found the corners inside the map, and the routine $\mathtt{CONNECT\_POLYGON}$ connects these corners in the correct order. In $\mathtt{BUILD\_GRAPH}$ we connect the vertices for the nearest neighborhoods and determine if an obstacle exist. The routine $\mathtt{PATH\_SEARCH}$ is responsible for finding a path within the graph.
• Figure 3: SkelUnet architecture. The scheme shows the different operations performed by each layer and how it affects the size and depth of the image. The arithmetic of these operations can be consulted in dumoulin2016guide.
• Figure 4: Comparative skeletonization results using Zhang-Suen's method versus SkelUnet. The first column is the input, the second column shows the output from Zhang-Suen's method, and the third column shows the SkelUnet output.
• Figure 5: Example of the environment for simulating drone dynamics. The quadrotor follows the paths obtained from SkelUnet in red, Zhang-Suen's method in green, and A* in yellow. This simulation validates dynamics; collision detection is performed using the dataset maps.