Robust 2D lidar-based SLAM in arboreal environments without IMU/GNSS

TL;DR

This work tackles SLAM in arboreal, GNSS-denied environments by deploying a minimalistic 2D lidar-based framework. It estimates the 2D pose $q_t=[x_t,y_t,theta_t]$ and builds a probabilistic map $M$ through scan matching using the Modified Hausdorff Distance (MHD), EKF pose refinement, and recursive map updates, all without IMU data. Key contributions include MHD-based scan matching without explicit data association, EKF fusion, and a robust occupancy-grid mapping strategy validated on real orchard datasets (CitrusFarm, Bacchus, Pullally) and controlled field trials, including deployment on a Unitree Go1. The method achieves sub-decimeter accuracy in controlled tests and demonstrates strong robustness in challenging outdoor conditions, often outperforming A-LOAM in GNSS-denied scenarios. Overall, the approach enables reliable autonomous navigation for precision agriculture using low-cost 2D lidar sensors and paves the way for future extensions to 6-DOF and 3D lidar integration.

Abstract

Simultaneous localization and mapping (SLAM) approaches for mobile robots remains challenging in forest or arboreal fruit farming environments, where tree canopies obstruct Global Navigation Satellite Systems (GNSS) signals. Unlike indoor settings, these agricultural environments possess additional challenges due to outdoor variables such as foliage motion and illumination variability. This paper proposes a solution based on 2D lidar measurements, which requires less processing and storage, and is more cost-effective, than approaches that employ 3D lidars. Utilizing the modified Hausdorff distance (MHD) metric, the method can solve the scan matching robustly and with high accuracy without needing sophisticated feature extraction. The method's robustness was validated using public datasets and considering various metrics, facilitating meaningful comparisons for future research. Comparative evaluations against state-of-the-art algorithms, particularly A-LOAM, show that the proposed approach achieves lower positional and angular errors while maintaining higher accuracy and resilience in GNSS-denied settings. This work contributes to the advancement of precision agriculture by enabling reliable and autonomous navigation in challenging outdoor environments.

Figures (8)

• Figure 1: Fruit orchard with dense canopies: RGB camera view (top) and 3D radar scan (bottom).
• Figure 2: Data filtering and scan matching with MHD.
• Figure 3: The mobile robot's pose $\bm{q}_t$ and lidar measurements $\bm{z}_{t,k}$ in an agricultural environment.
• Figure 4: Slicing of 3D lidar measurements and 2D projection onto "ground" plane.
• Figure 5: Online SLAM navigation based on scan matching.