Gaussian Process-based Traversability Analysis for Terrain Mapless Navigation

TL;DR

This work proposes a new geometric-based uneven terrain mapless navigation framework combining a Sparse Gaussian Process local map with a Rapidly-Exploring Random Tree* (RRT*) planner, which enables ground robots to safely navigate environments with varying elevations and steep obstacles.

Abstract

Efficient navigation through uneven terrain remains a challenging endeavor for autonomous robots. We propose a new geometric-based uneven terrain mapless navigation framework combining a Sparse Gaussian Process (SGP) local map with a Rapidly-Exploring Random Tree* (RRT*) planner. Our approach begins with the generation of a high-resolution SGP local map, providing an interpolated representation of the robot's immediate environment. This map captures crucial environmental variations, including height, uncertainties, and slope characteristics. Subsequently, we construct a traversability map based on the SGP representation to guide our planning process. The RRT* planner efficiently generates real-time navigation paths, avoiding untraversable terrain in pursuit of the goal. This combination of SGP-based terrain interpretation and RRT* planning enables ground robots to safely navigate environments with varying elevations and steep obstacles. We evaluate the performance of our proposed approach through robust simulation testing, highlighting its effectiveness in achieving safe and efficient navigation compared to existing methods.

Figures (6)

• Figure 1: Overview of the proposed uneven terrain navigation framework. From left to right: Localization and LiDAR PointCloud data informs the Gaussian Process Mapping Module. An RBF Kernel trains a model on the induced PointCloud points within this module, yielding Elevation, Variance, and Gradient local cost maps. These maps feed into the Path Planning Module, which constructs a traversability local cost map based on terrain attributes. RRT* planning occurs within the local traversability map and plans a path to the goal. During planning optimizations, including a footprint-based assessment approach and local sub-goal selection. Finally, we send waypoints to the Differential Drive Controller, which sends control commands to the autonomous vehicle.
• Figure 2: Traversability analysis for one observation (marked by a black square in Fig. \ref{['fig_sim_d']}) in environment B: (a) shows the simulated environment. (b), (c), and (d) show the local elevation $\mathcal{M}_h$, slope $\mathcal{M}_{\Delta}$, and uncertainty maps $\mathcal{M}_{\sigma}$ respectively, generated by the SGP elevation model. (e) and (f) visualize the flatness $f$ and step height $\zeta$ maps. (g) shows the final local traversability map $\mathcal{M}_{\tau}$ with the uncertainty map $\mathcal{M}_{\sigma}$ applied as a mask. (h) shows the generated RRT* planning on top of the final traversability map. The traversability values are presented on a rainbow spectrum, with purple representing traversable and red representing non-traversable. Similarly, the Uncertainty map is represented on a grey color gradient where white represents uncertain and black indicating certain regions.
• Figure 3: Figure (a) is the Gazebo world for Environment A, provided by jian2022putn, and (b) is the associated topographic map with the safest path achieved by each tested algorithm for task $T_1$. Figure (c) is the Gazebo world for Environment B, and beside it, (d) depicts the topographic map with overlaid paths for all trials run in Tasks 3 and 4. The black box represents the local map $\mathcal{M}$ view, which is visualized in Fig. \ref{['fig_trav']}.
• Figure 4: Expanded Fig. \ref{['fig_trav_h']} illustrates an RRT* planning iteration, where the green trajectory highlights the route to the chosen sub-goal. Black nodes signify edge nodes leading the untraversable areas. While green nodes indicate frontier nodes and potential targets for subsequent planning iterations.
• Figure 5: Environment A velocity graphs. The left graph displays the velocities for 10 trials for each method tested in $T_1$. AL and GT stand for ALOAM and Ground Truth respectively. The right graph displays a single velocity for the safest path taken displayed in Fig. \ref{['fig_sim_b']}.