cs.ROAug 27, 2026

One year in a forest: Analyzing the challenges of autonomous navigation in subarctic environments

Authors: Matěj Boxan, Nicolas Lauzon, Veronica Vannini, Mathis Turgeon-Roy, François Pomerleau

Organizations: Norlab, Université Laval, Quebec, Canada

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Subarctic regions have the potential to see increased deployment of autonomous robots in applications including forestry, mining, and environmental monitoring. In these conditions, an autonomous system's reliance on GNSS or cloud computing is precarious due to dense tree canopies and atmospheric attenuation, necessitating onboard sensing and data processing. However, established exteroceptive modalities, including cameras, lidars, and radars, are typically evaluated in structured urban settings or in environments that lack significant seasonal variations. To address this, we present a field report on a year-long deployment of a mobile robot in a subarctic boreal forest. We evaluate 64 km of data using nine odometry, localization, and mapping methods and assess their performance across seasonal changes. The performed experiments suggest that the environment changes significantly hinder the performance of state-of-the-art techniques, which show increased fragility when subject to conditions characterized by self-similar scenes or tall snowbanks. Additionally, complex Simultaneous Localization and Mapping (SLAM) algorithms offer limited accuracy gains over a proprioceptive baseline while significantly increasing system fragility. Furthermore, by correlating the position drift with features and confidence weight distribution, we show that visual-based SLAM methods are particularly affected by the seasonal changes. Additionally, we investigate the task of cross-season localization in a prior map. While lidar-based methods successfully completed localization runs between seasons, radar and visual methods are prone to failure due to a few matching features between runs, even within the same season. Finally, we detail the challenges and lessons learned from this year-long trial, including a multi-season Teach and Repeat (T&R) evaluation using both radar and lidar-based pipelines.

Figures & tables

Figure 1: Example of seasonal changes in subarctic climate affecting a boreal forest. These conditions are used to evaluate the impact of seasonal variations on the performance of odometry, localization, and mapping algorithms. To this extent, we study the effects of seasonal changes on radar, lidar, and visual-based methods, based on a field campaign spanning one year.
 CodeClimate NameCriteria
temperateCfaHumid subtropicalHot humid summer, mild winter
CfbTemperate oceanicWarm summer, mild winter
CsaHot-summer MediterraneanHot dry summer, mild winter
continentalDfaHumid continentalHot summer, 1 month above 22 \SIUnitSymbolCelsius22\text{\,}\mathrm{\SIUnitSymbolCelsius}
DfbHumid continentalWarm summer
DfcSubarcticColdest month below −3 \SIUnitSymbolCelsius-3\text{\,}\mathrm{\SIUnitSymbolCelsius}
Table 1: Köppen-Geiger climate classification for a subset of zones where robots were deployed over multiple seasons. The Subarctic climate zone is the focus of this paper.
Figure 2: Distribution of data recordings and localization and mapping methods that include seasonal variations across the globe, with their Köppen-Geiger climate classification. For simplicity, we do not include climate zones not covered by multi-season datasets or not present in the evaluation of localization and mapping methods in unstructured environments. The map was cropped to focus on the identified deployments. Image inset were adapted from [ 56 , 60 , 29 , 40 ] .
Figure 3: Illustration of the main differences between evaluation focusing on parameters as opposed to systems as a whole. Left : Parameter evaluations highlight behaviors in controlled environments. Right : Evaluation targeting system behaviors in an uncontrolled environment where disturbances (disturb.) and stresses will influence the localization errors, which are harder to control.
PerturbationTypePotential cause
CollisionsDDriver inattention; vegetation blocking path
Hardware failureDCold temperature; wear and tear; accidents
Software failureDInteger overflow; memory leaks
Rain/precipitationS/DHumid climate zones
Wheel slippageS/DIce, snow; dirt; wet and soft terrain
Env. changesSLocation of the task or mission
Table 2: Examples of perturbations (i.e., stressors and disturbances) observed in our field deployments.
Figure 4: Example taxonomy of localization and mapping algorithms in multi-season outdoor environments. The figure depicts the expected behavior of each method, where the method’s deviation from the GT decreases with raising complexity. Green, orange, and blue lines originate from the same data, while the red line represents localization at a later date.
Figure 5: An illustration of the discussed error metrics. RTE ( RTE ) requires orientations to correctly align the two sub-trajectories. SARTE ( SARTE ) uses part of the previous window as a proxy to estimate the orientation. RTDE ( RTDE ) is distance-based and does not account for errors in orientation. Numerical values in this toy example correspond the red lines lengths.
Figure 6: Elevation map of the deployment site with the six evaluated trajectories. The bottom-right inset shows the location of the site within the province of Quebec, Canada.
YearMethodRef.Wheel encodersIMUStereo cam.LidarRadarLCLoc. modeDim
2011Proprioceptive[ 54 ]  –––––3D
2022WILN[ 4 ]  – –– 3D
20252Fast-2Lamaa[ 47 ]– – –  3D
2021ORB-SLAM3[ 13 ]–  ––  3D
2025cuVSLAM[ 41 ]–– ––  3D
2021DROID-SLAM[ 76 ]–– –– –3D
Table 3: Summary of the evaluated methods.
Figure 7: Reported mean SARTE ( SARTE ) and improvement or degradation over Proprioceptive as percentage points (  %pt\text{\,}\mathrm{\%pt} ) for eleven evaluated methods over six trajectories in a subarctic environment. The failure rate of each method follows the method’s name using (X %). The first column shows the total degradation of a method type (Exteroceptive odometry, Exteroceptive odometry with PGO ) over Proprioceptive. The second column shows the total degradation of each method over Proprioceptive. The remaining values reported for each trajectory are averaged over all 12 recordings throughout the year. A star (*) denotes the odometry output of a method.
 2Fast-2LamaaORB-SLAM3cuVSLAMKISS-SLAMDROID-SLAMNR-SLAM
Red (12)-0.4 (0)-0.5 (9)0.0 (0)0.0 (0)0.0 (242)-3.6 (14)
Blue (9)-0.1 (0)-5.8 (8)0.0 (0)0.0 (0)-0.1 (600)52.2 (1.1k)
Green (21)-0.3 (1)-6.3 (10)0.0 (0)0.0 (0)0.2 (4.7k)54.3 (1.5k)
Orange (9)-0.1 (0)-45.5 (3)0.0 (1)0.4 (7)1.2 (22k)83 (3.7k)
Yellow (19)-0.2 (3)-10.7 (9)0.0 (6)0.0 (3)0.0 (22k)75.5 (2.8k)
Magenta (0/7)-0.1 (18)-2.9 (3)0.0 (5)0.0 (0)1.8 (3.3k)4.0 (656)
Table 4: Loop closure statistics across trajectories. Reported improvement or degradation as percentage points (  %pt\text{\,}\mathrm{\%pt} ) of SARTE over the odometry counterparts for six evaluated methods. The number in parentheses after the performance change indicates the number of detected loop closures. The number of expected loop closures for each trajectory is presented in parentheses in the first column. For Magenta, the dash differentiates between the expected loop closure count for range sensors and visual methods.
Figure 8: A comparison of RTDE ( RTDE ) performance of proprioceptive, lidar, radar, and visual-based odometry algorithms across seasons.
Figure 9: Effect of seasonal changes on ORB and cuVSLAM feature detection and DROID-SLAM network confidence. The top row shows ZED X camera images taken in the Orange trajectory in summer and winter. The second and third rows displays heatmaps of the accumulated ORB and cuVSLAM features. Bottom row shows a heatmap of accumulated DROID-SLAM confidence weights. The data originates from a ten-second window preceding the top row image capture. The color-maps are normalized for each individual image.
Figure 10: Relation between RTDE ( RTDE ) and the distribution of ORB and cuVSLAM features and DROID-SLAM weights across seasons. IQR stands for IQR . Lower number of detected features leads to higher error values and correlates to images originating in winter months. Note that the scale of RTDE differs between the three graphs.
MethodMetricSummerWinter
ORB-SLAM3Bottom half56 %56\text{\,}\mathrm{\%}1.8 %1.8\text{\,}\mathrm{\%}
(ORB features)RTDE min0.74 %0.74\text{\,}\mathrm{\%}1.31 %1.31\text{\,}\mathrm{\%}
 RTDE max0.99 %0.99\text{\,}\mathrm{\%}2.47 %2.47\text{\,}\mathrm{\%}
cuVSLAMBottom half74.5 %74.5\text{\,}\mathrm{\%}0.8 %0.8\text{\,}\mathrm{\%}
(Shi-Tomasi keypoints)RTDE min1.30 %1.30\text{\,}\mathrm{\%}1.90 %1.90\text{\,}\mathrm{\%}
 RTDE max1.85 %1.85\text{\,}\mathrm{\%}7.35 %7.35\text{\,}\mathrm{\%}
Table 5: Effects of seasonal changes on visual-based methods. Bottom half denotes the percentage of features or confidence weights in the lower half of the image. RTDE limits report the range of per-bin median RTDE , computed over bins where at least 75 %75\text{\,}\mathrm{\%} of frames come from the respective season.
Figure 11: Evaluation of cross-seasonal localization performance on combined data from Orange and Yellow trajectories. Light cells mean better performance. Each row indicates the season and month used to construct the prior offline map. The Control columns shows localization performance on the same data as the mapping sequence. Each remaining column denotes the season and month of the active localization run. The LFR ( LFR ), displayed in the top triangle of each cell, is calculated relative to the full trajectory duration. We omit the ratio when the localization failed within the first 100 m100\text{\,}\mathrm{m} (i.e., black cells with N/A) without recovery, or when the full trajectory was completed successfully. SARTE stands for SARTE . Because of its architectural limitation, cuVSLAM was unable to successfully initialize outside of the three control sequences.
Figure 12: Challenges and lessons learned encountered during our year-long field campaign: (a, b) Mobility in snow is related not only to its depth, but also to other properties. (a) Shows the Warthog UGV in the end of November, easily navigating over 40 cm40\text{\,}\mathrm{cm} of fresh snow. In April (b), the robot’s wheels were freely spinning as the wet spring snow didn’t provide enough friction for the differential-drive robot to move forward. (c) Any permanently installed equipment, such as the depicted GNSS antenna, must be cleaned regularly or equipped with heating elements to remove snow cover, which effectively blocks signals. (d) Performing a 6- DoF manipulation with a half-a-tonne robot for extrinsic calibration requires special equipment. We used an industrial crane to excite all axes of the IMU .
Figure 13: Sensor data comparison between seasons. First row shows images from the left lens of the ZED X camera. Second row displays radar submaps with the robot’s trajectory in orange and snowbanks highlighted in blue. The last row depicts a part of the global lidar point cloud, again with the robot’s trajectory in orange. Green indicates height and vegetation, while the blue color in the winter column highlights snow accumulation around the road. Yellow spheres in the orange trajectories indicate the UGV ’s position where the corresponding image in the top row was taken.
Figure 14: Qualitative results for lidar odometry methods and the Proprioceptive baseline.
Figure 15: Qualitative results for lidar odometry methods with PGO and the Proprioceptive baseline.
Figure 16: Qualitative results for radar odometry methods and the Proprioceptive baseline.
Figure 17: Qualitative results for lidar odometry methods with PGO and the Proprioceptive baseline.
Figure 18: Qualitative results for visual odometry methods and the Proprioceptive baseline.
Figure 19: Qualitative results for visual odometry methods with PGO and the Proprioceptive baseline.
More in this topic

Off-Road Robot Navigation

Off-road robot navigation studies how ground robots perceive terrain, assess traversability, and plan safe routes through unstructured environments.

See all 54 papers

Explore similar work

CardsList
Mar 9, 2026

FoMo: A Multi-Season Dataset for Robot Navigation in Forêt Montmorency

The Forêt Montmorency (FoMo) dataset is a comprehensive multi-season data collection, recorded over the span of one year in a boreal forest. Featuring a unique combination of on- and off-pavement environments with significant environmental changes, the dataset challenges established odometry and SLAM pipelines. Some highlights of the data include the accumulation of snow exceeding 1 m, significant vegetation growth in front of sensors, and operations at the traction limits of the platform. In total, the FoMo dataset includes over 64 km of six diverse trajectories, repeated during 12 deployments throughout the year. The dataset features data from one rotating and one hybrid solid-state lidar, a Frequency Modulated Continuous Wave (FMCW) radar, full-HD images from a stereo camera and a wide lens monocular camera, as well as data from two IMUs. Ground Truth is calculated by post-processing three GNSS receivers mounted on the Uncrewed Ground Vehicle (UGV) and a static GNSS base station. Additional metadata, such as one measurement per minute from an on-site weather station, camera calibration intrinsics, and vehicle power consumption, is available for all sequences. To highlight the relevance of the dataset, we performed a preliminary evaluation of the robustness of a lidar-inertial, radar-gyro, and a visual-inertial localization and mapping techniques to seasonal changes. We show that seasonal changes have serious effects on the re-localization capabilities of the state-of-the-art methods. The dataset and development kit are available at https://fomo.norlab.ulaval.ca.

Matěj Boxan, Gabriel Jeanson, Alexander Krawciw +5

Norlab, Universit´e Laval, Quebec City, Canada · University of Toronto Robotics Institute, Toronto, Canada

Simultaneous Localization and MappingRobotic Data CollectionRobot NavigationLiDAR OdometryRobot Localization
AskPDF
Apr 21, 2026

Radar Odometry Subject to High Tilt Dynamics of Subarctic Environments

Rotating FMCW radar odometry methods often assume flat ground conditions. While this assumption is sufficient in many scenarios, including urban environments or flat mining setups, the highly dynamic terrain of subarctic environments poses a challenge to standard feature extraction and state estimation techniques. This paper benchmarks three existing radar odometry methods under demanding conditions, exhibiting up to 13° in pitch and 4° in roll difference between consecutive scans, with absolute pitch and roll reaching 30° and 8°, respectively. Furthermore, we propose a novel radar-inertial odometry method utilizing tilt-proximity submap search and a hard threshold for vertical displacement between scan points and the estimated axis of rotation. Experimental results demonstrate a state-of-the-art performance of our method on an urban baseline and a 0.3% improvement over the second-best comparative method on a 2-kilometer-long dynamic trajectory. Finally, we analyze the performance of the four evaluated methods on a complex radar sequence characterized by high lateral slip and a steep ditch traversal.

Matěj Boxan, William Larrivée-Hardy, François Pomerleau

Northern Robotics Laboratory, Université Laval, Quebec, Canada. · Université Laval, Quebec, Canada.

Robot NavigationRadar-Inertial OdometryRobot Localization
AskPDF
Mar 9, 2026

Perception-Aware Communication-Free Multi-UAV Coordination in the Wild

We present a communication-free method for safe multi-robot coordination in complex environments such as forests with dense canopy cover, where GNSS is unavailable. Our approach relies on an onboard anisotropic 3D LiDAR sensor used for SLAM as well as for detecting obstacles and neighboring robots. We develop a novel perception-aware 3D navigation framework that enables robots to safely and effectively progress toward a goal region despite limited sensor field-of-view. The approach is evaluated through extensive simulations across diverse scenarios and validated in real-world field experiments, demonstrating its scalability, robustness, and reliability.

Manuel Boldrer, Michal Kamler, Afzal Ahmad +1

Department of Cybernetics, Czech Technical University in Prague, Karlovo namesti 13, 121 35 Prague 2, Czechia.

Multi-Robot CoordinationUAV NavigationSafe Robot NavigationMulti-Robot Motion Planning
AskPDF