Online Adaptive Traversability Estimation through Interaction for Unstructured, Densely Vegetated Environments

TL;DR

This paper tackles the challenge of robust traversability estimation for autonomous ground vehicles in unstructured, densely vegetated environments by introducing a lidar-only online adaptive TE method. It fuses robot experience with a 3D probabilistic voxel map into a sparse graph (ograph) and updates a UNet-based traversability estimator in real time, enabling self-supervised learning directly on the robot. Key contributions include the online data generation and graph-based data fusion, the voxel-layer feature representation, and a costmap synthesis pipeline that supports safe path planning under online adaptation. Experiments in forested and industrial settings show that the approach achieves MCC scores around 0.63–0.71, with safe navigation demonstrated in real-time in vegetation, using limited computational resources. The work provides practical guidance on training strategies for online adaptation and offers open-source resources to foster further development in lidar-based online TE for field robotics.

Abstract

Navigating densely vegetated environments poses significant challenges for autonomous ground vehicles. Learning-based systems typically use prior and in-situ data to predict terrain traversability but often degrade in performance when encountering out-of-distribution elements caused by rapid environmental changes or novel conditions. This paper presents a novel, lidar-only, online adaptive traversability estimation (TE) method that trains a model directly on the robot using self-supervised data collected through robot-environment interaction. The proposed approach utilises a probabilistic 3D voxel representation to integrate lidar measurements and robot experience, creating a salient environmental model. To ensure computational efficiency, a sparse graph-based representation is employed to update temporarily evolving voxel distributions. Extensive experiments with an unmanned ground vehicle in natural terrain demonstrate that the system adapts to complex environments with as little as 8 minutes of operational data, achieving a Matthews Correlation Coefficient (MCC) score of 0.63 and enabling safe navigation in densely vegetated environments. This work examines different training strategies for voxel-based TE methods and offers recommendations for training strategies to improve adaptability. The proposed method is validated on a robotic platform with limited computational resources (25W GPU), achieving accuracy comparable to offline-trained models while maintaining reliable performance across varied environments.

Figures (9)

• Figure 1: The overview of the online TE adaptation method. The ForestTrav method is augmented with new modules to facilitate online learning. The online node generates and trains the new model from the self-supervised labelled data collected online, stored as a graph representation. A fusion module generates the data, combining the collision map and 3D probabilistic environmental representation at fixed-time intervals into the ograph.
• Figure 2: Visualisation of the Learning from Experience of the robot. The top image shows the robot in a real-world scenario, with green for the collision-free state and red for the collision state. The middle image shows the robot labelling the 3D probabilistic map as collision-free, ellipsoid representing the NDT representation of a voxel. The bottom image shows the robot in collision, labelling the voxels as non-traversable in the bound box at the front of the robot, with the bounding volume extending a short distance beyond the robot chassis.
• Figure 3: The Left image shows a 3D traversability map, and the right image shows a costmap with traversable (green), non-traversable (red) and virtual surfaces. The traversable elements can have low or high costs. Virtual cells are cells that have not been directly observed, and the cost is estimated based on their surroundings.
• Figure 4: Top left: Shows the overview of the environment with the orange dots showing the locations of data sets from ForestTrav ruetz2024foresttrav. The blue dots are the novel data set introduced in this work from an industrial environment. An example is shown in the top right. The bottom left shows an example of the SPARSE environment, the bottom right scene from a densely vegetated forest.
• Figure 5: Comparison of two ensembles of models trained on either the dense data set in the top row (A, B, C) or the industrial data set in the bottom row (D, E, F). The left column shows each model's TE in an industrial environment, the middle and right columns show the TE of two densely vegetated environments with a variation of tree sizes and underbrush.