Off‐road ground robot path energy cost prediction through probabilistic spatial mapping

Abstract

Energy is an important factor in planning for ground robots, particularly in off‐road environments where the terrain significantly impacts energy usage. Unfortunately, energy costs in these environments are variable and uncertain, making planning difficult. In this paper, we present a method, based on Gaussian process regression (GPR) and known vehicle modeling information, for predicting future path energy costs. The method uses data, collected by a robot operating in a 3D environment with varying terrains, to build a map from inputs (including position, heading, slope, and satellite imagery) to power consumption. The energy costs of future paths are predicted through the summation of power predictions. Importantly, correlations in those predictions are considered to avoid overconfidence. Experimental cross‐validation results demonstrate improved accuracy of path energy cost predictions against a baseline approach, as well as the effect of Gaussian process inputs and kernel choice. Additionally, we show how vehicle modeling can aid in predicting energy costs, particularly when data on the environment is sparse.

Learning-Based Methods of Perception and Navigation for Ground Vehicles in Unstructured Environments: A Review

“…Although beyond of the scope of this review, we can easily perceive how the need to combine multiple heterogeneous platforms especially with potentially different perspectives, is still strong in the context of rough environment navigation. Some works suggesting the adoption of overhead or satellite imagery have been presented [ 36 , 41 , 80 ]; others have used terrain height maps [ 25 , 29 , 57 ]. However, for some of them, it could easily be figured out that such images or elevation maps could be provided by an aerial vehicle hovering and following the ground vehicle.…”

“…An off-road navigation strategy is presented by Quann et al [ 36 ] based on probabilistic energy cost prediction. A Gaussian process regressor realizes a mapping from current robot pose, terrain slope (along robot motion direction) and grayscale satellite imagery to power consumption.…”

An autonomous navigation approach for unmanned vehicle in outdoor unstructured terrain with dynamic and negative obstacles

“…But he maps the 3D point clouds directly into a 2D grid, making them lose terrain features. The probabilistic energy cost map is proposed by Quann et al to measure the traversal cost [19], which is built from the pose of robot, the terrain slope, and the satellite imagery. However, this method relies heavily on the vehicle modeling information, which makes it difficult to perform in different type of vehicles.…”

Autonomous robotic exploration with simultaneous environment and traversability models learning

“… Sofman et al (2006) incrementally learned the relation between dense laser-based features characterizing ground unit traversability and overhead features that can be used to assess traversability from aerial images, whereas Bekhti and Kobayashi (2016) learned to predict vibration-based traversability from terrain texture. Quann et al (2020) proposed an energy traversal cost regressor considering both terrain position and appearance. In addition, Mayuku et al (2021) proposed a self-supervised labeling approach for a near-to-far scenario, where vibration-based traversal cost is inferred from image data, and the self-supervised data gathering is based on identified terrain classes.…”