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

arXiv:2603.08433 · cs.RO · Submitted 2026-03-09 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: I'm Rosa, and with me are Dev and Taro, guest researcher.

Dev: Today's paper: "FoMo: A Multi-Season Dataset for Robot Navigation in For et Montmorency".

Rosa: The Foret Montmorency (FoMo) dataset is a comprehensive, multi-season data collection recorded over one year in a boreal forest, featuring unique environmental challenges like significant snow accumulation and evolving terrain.

Dev: First, who's behind it and why it matters.

Title and authors: Rosa: Now, let's move into what the actual summary of "FoMo: A Multi-Season Dataset for Robot Navigation in For et Montmorency" is actually saying about the data collection process and what makes it unique.

Dev: They are highlighting that the primary value of this collection lies in capturing significant environmental changes, specifically mentioning snow accumulation exceeding one meter and substantial vegetation growth right in front of the sensors.

Taro: That environmental variability is what really pushes the limits for localization algorithms; it’s not just static noise, but dynamic physical obstructions changing constantly.

Rosa: And they emphasize that the dataset documents how terrain traversability evolves throughout the seasons, showing a platform getting stuck in mud pits that were frozen in winter.

Dev: This highlights a major challenge for any navigation system: adapting to conditions that are fundamentally different from what was encountered during initial training or calibration.

Taro: It seems like the authors are providing concrete examples of failure modes caused by these seasonal shifts, which is exactly what we need to build better resilience into our autonomy.

Rosa: They also detail the sensor suite they used, including two lidars—one rotating and one hybrid solid-state—along with a Frequency Modulated Continuous Wave radar and full-HD stereo cameras.

Dev: That multi-modal approach is important because it means the data captures different types of environmental cues simultaneously, which should help in distinguishing between snow and actual ground features.

Taro: Having both Lidar and Radar on board gives the system redundancy when one sensor might be temporarily blinded by heavy snow or dense foliage.

Rosa: So, in short, they've packaged a year of complex boreal forest navigation data with a rich set of sensors to create a dataset specifically designed to challenge current localization methods.

The paper's summary: Dev: Moving on to the suggested improvements within the paper for dealing with these challenges, they focus heavily on developing localization algorithms that can explicitly model and adapt to those non-linear environmental changes.

Rosa: They suggest integrating sensor fusion frameworks, like Lidar-Inertial Odometry combined with learned or adaptive motion models that account for external variables like snow accumulation.

Taro: That points toward the need for AI systems to have motion models that aren't just static equations but can dynamically adjust based on what the sensors are currently reporting about their surroundings.

Dev: The paper also suggests training or fine-tuning these fusion frameworks specifically on multi-modal data that has a high dynamic range in sensor readings, like the difference between snow and terrain features.

Rosa: This means we need AI systems capable of handling that kind of sensory noise and variation without losing track of their position during severe weather events.

Taro: I think this ties into our work on semantic terrain understanding; if the system can perceive *what* it is encountering—snow, mud, or rock—it can make better decisions about movement.

Dev: That aligns with the idea of using deep learning models, perhaps Graph Neural Networks or Transformer architectures, to fuse sparse point cloud data from Lidar and Radar with dense visual features from the cameras.

Rosa: If we can generate a semantically rich three dee map representation that updates in real-time based on seasonal surface types, the robot can perform better path planning and traversability estimation <ref:2603.08433#pg2>.

The paper's improvements: Rosa: So, to wrap up the discussion on "FoMo: A Multi-Season Dataset for Robot Navigation in For et Montmorency," it really comes down to how this collection pushes us toward more robust and adaptive navigation systems.

Dev: The main implication is that we need localization algorithms that are not just tuned for specific conditions but can handle the kind of extreme, non-linear environmental shifts documented in this data.

Taro: The impact could be significant because it gives engineers a way to validate if their autonomy systems can survive prolonged missions where conditions change drastically over a single year.

Rosa: And the dataset itself provides the necessary real-world complexity to ensure that when we deploy these systems, they have already encountered a wide range of difficult scenarios.

Dev: The authors clearly show that success hinges on integrating sensor fusion with models that can handle high levels of environmental variance, which is something we need to focus on at the control level.

Taro: I think the ability to handle unpredictable world behavior, like getting immobilized in a frozen mud pit described in their summary, is what really matters for future autonomy.

Rosa: Absolutely, and as we look ahead at this paper from "FoMo: A Multi-Season Dataset for Robot Navigation in For et Montmorency," it sets a high standard for creating data that truly stresses the limits of current SLAM techniques.

Dev: It gives us a clear direction on what to prioritize when designing systems that need long-term reliability in unpredictable outdoor settings.

Taro: It’s an excellent resource for pushing the boundaries of what we think is achievable in autonomous navigation under severe environmental stress.

Conclusion: Rosa: So, to wrap up our discussion on "FoMo: A Multi-Season Dataset for Robot Navigation in For et Montmorency," we've seen how this collection provides a truly comprehensive testbed for challenging localization and mapping systems across diverse and extreme conditions.

Dev: Exactly, the sheer scale of the environmental variability documented here really puts our current state estimation techniques to the test, especially concerning those long-term drift issues we see in complex scenarios.

Rosa: I think what stands out is how they’ve meticulously structured their data to cover everything from pure off-trail navigation in dense vegetation to severe snow accumulation that can exceed a meter.

Taro: That multi-season aspect is crucial because it forces any autonomous system to deal with terrain traversability evolving over time, not just static obstacles.

Dev: The sensor suite they used, combining Lidar and Radar with stereo vision, gives us a fantastic foundation for testing how well different modalities can compensate for sensor occlusion caused by snow or dense forest growth.

Rosa: It’s exciting to think about the potential impact this has on real-world applications; if we can train systems on data this rich, they could handle much more unpredictable outdoor environments than we currently allow them to.

Taro: I agree, and what I found particularly interesting was how they've defined the Ground Truth using a multi-step optimization process involving GNSS receivers at different locations.

Dev: That rigorous GT generation is what makes this dataset so valuable for benchmarking; it gives us a solid, verifiable baseline against which we can measure how much better our loop closure or re-localization methods are performing.

Rosa: We should definitely keep an eye on this type of comprehensive data collection going forward because it sets a very high bar for creating truly representative datasets.

Taro: Indeed, and the implications for safety in autonomous systems are huge if these localization challenges can be solved reliably across varying conditions.

Dev: It’s definitely something we need to keep our eyes on as we push the limits of real-time performance and failure mode analysis in control systems.

Rosa: Well, that’s all the time we have for this session on "FoMo: A Multi-Season Dataset for Robot Navigation in For et Montmorency." Next up, we're going to look at how recent VLA models are handling long-horizon planning with ProbeFlow.

Norlab Universitˇe Laval

cs.RO

Submitted: 2026-03-09

Updated: 2026-10-05

DOI: 10.1177/02783649261491481

Code: https://github.com/ripl-lab/allan_variance_ros2

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 76/100

The gist: The Foret Montmorency (FoMo) dataset is a comprehensive, multi-season data collection recorded over one year in a boreal forest, featuring unique environmental challenges like significant snow

Key concepts

Multi-Season Data Collection: The dataset spans over a full year, capturing how the boreal forest environment changes dramatically between winter and summer. This variability includes temperature shifts from -20°C to 20°C and snow depths reaching up to 120 cm, providing a realistic test for navigation systems that must handle changing conditions.
Diverse Trajectory Types: The dataset includes six distinct paths designed to test different navigation challenges. These range from short, on-road baselines to long, off-trail routes through dense vegetation and rocky terrain. This diversity ensures the evaluation covers a wide spectrum of mobility constraints and environmental complexities.
Exhaustive Sensor Suite: The data incorporates a rich set of sensors including two lidars, an FMCW radar, stereo and monocular cameras, and two IMUs. This comprehensive sensor package provides multiple modalities for localization, allowing researchers to evaluate how different sensor types perform when facing challenges like sensor occlusion or strong vibrations.
Ground Truth Generation: The Ground Truth (GT) is created by meticulously post-processing GNSS receiver data from three mounted receivers and a static base station. This multi-step optimization process yields accurate trajectory points, which serve as the objective standard against which localization and mapping algorithms are evaluated.

Terminology

Calibration — 29× in this paper · explained in 7 episodes
Ground Truth — 15× in this paper
Benchmark — 8× in this paper · explained in 2 episodes
Robustness — 6× in this paper · explained in 5 episodes
Frequency Modulated Continuous Wave — 5× in this paper
Benchmarking — 4× in this paper · explained in 3 episodes
Point Clouds — 4× in this paper · explained in 2 episodes
Uncrewed Ground Vehicle — 4× in this paper
Alignment — 2× in this paper · explained in 1 episode
Core — 2× in this paper · explained in 2 episodes
International Conference on Robotics and Automation — 2× in this paper
Multimodal Dataset — 2× in this paper · explained in 2 episodes
Precision Time Protocol — 2× in this paper
Accuracy — 1× in this paper · explained in 1 episode
Adam — 1× in this paper · explained in 2 episodes
Bacchus Long-Term — 1× in this paper
Continuously Operating Reference Station — 1× in this paper
Data Fusion — 1× in this paper · explained in 3 episodes
Decode — 1× in this paper · explained in 2 episodes
Failure Modes — 1× in this paper · explained in 3 episodes
Fine-Tuning — 1× in this paper · explained in 5 episodes
Global Navigation Satellite System — 1× in this paper
Iterative Closest Point — 1× in this paper
Point Cloud Data — 1× in this paper · explained in 2 episodes
Real Time Kinematic — 1× in this paper
Redundancy — 1× in this paper · explained in 2 episodes
Self-Supervised Learning — 1× in this paper · explained in 8 episodes
Transformer — 1× in this paper · explained in 2 episodes
Transformer Architecture — 1× in this paper · explained in 2 episodes
VLA Models — 1× in this paper · explained in 3 episodes

Summary

The Foret Montmorency (FoMo) dataset is a comprehensive, multi-season data collection recorded over one year in a boreal forest, featuring unique environmental challenges like significant snow accumulation and evolving terrain. This dataset is crucial for challenging established odometry and SLAM pipelines by providing data spanning multiple seasons and diverse conditions, which are known to seriously affect the re-localization capabilities of state-of-the-art methods.

The gist

The FoMo dataset is a comprehensive, multi-season data collection recorded over one year in a boreal forest featuring unique environmental challenges like significant snow accumulation and evolving terrain.

Dataset Composition and Scope

The FoMo dataset spans over one year and includes 64 km of six diverse trajectories, repeated during 12 deployments throughout the year. It features a unique combination of on-road, off-pavement, and off-trail conditions, including the vehicle’s high pitch and roll angles, strong vibrations, and sensor occlusion. The data collection took place between November 2024 and October 2025 in Foret Montmorency at an altitude of 600 m.

The dataset incorporates an exhaustive sensor suite:

Two lidars (one rotating and one hybrid solid-state).
A Frequency Modulated Continuous Wave radar (FMCW).
Full-HD images from a stereo camera and a wide lens monocular camera.
Two Inertial Measurement Units (IMUs).

Additional metadata is available for all sequences, including one measurement per minute from an on-site weather station, camera calibration intrinsics, and vehicle power consumption. Ground Truth (GT) is calculated by post-processing three GNSS receivers mounted on the Uncrewed Ground Vehicle (UGV) and a static GNSS base station.

Environmental Variability and Challenges

A core feature of the dataset is substantial seasonal changes in the boreal forest environment. Key environmental highlights include:

snow accumulation exceeding 1 m

significant vegetation growth in front of sensors

The environment offers variability across a range of temperatures, with data covering conditions from −20 °C to 20 °C and featuring snow on the ground for five months of the year. The data collection documents how terrain traversability evolves from winter to summer months, with the platform becoming immobilized in a mud pit that was originally frozen in winter.

The dataset includes detailed measurements from an onsite meteorological weather station covering the full span of the data recording at a rate of one reading per minute. Figure 3 depicts the snow height and ambient air temperature over a year, showing conditions ranging from −20 °C to 20 °C and snow depths up to 120 cm.

Trajectory Diversity

The six recorded trajectories are color-coded (Yellow, Red, Blue, Orange, Green, and Magenta) and include various surface types and environmental complexities. The trajectories were chosen with respect to the platform’s mobility and constraints from the chosen GT method. Key trajectory characteristics include:

The shortest trajectory (Red), 300 m long, acts as an on-road baseline confined to paved surfaces.
Two trajectories (Green and Magenta), both 700 m long, are dedicated to challenging, pure off-trail navigation (Green characterized by steep ascent/descent and dense vegetation; Magenta focuses on highly irregular, rocky terrain like a stone quarry).
The three remaining trajectories feature mixed environments: Blue (550 m) and Yellow (1900 m) combine on-road with off-pavement gravel road segments, often involving long uphills and significant seasonal challenges like deep snow or snowbanks.
The longest path, Orange (2200 m), also mixes on-road and off-pavement.

Ground Truth Generation

The Ground Truth (GT) trajectories are generated by a multi-step pipeline for each trajectory of each deployment:

Post-process the recorded GNSS Receiver Independent Exchange Format (RINEX) data using PPK in Emlid Studio to correct the positions of the static Emlid Reach RS3 receiver using a Continuously Operating Reference Station (CORS) located in Quebec City.
Compute trajectories for each of the three GNSS receivers where t represents time, yielding points p k(t) and covariance matrices W k(t).
Combine these three GNSS trajectories into a single trajectory τ̂(t) = ˆp t using an optimization problem based on the point-to-Gaussian distance metric (Equation 3), utilizing the known set of three non-collinear points representing the antenna locations.

Evaluation and Benchmarking

The evaluation focuses on localization and mapping tasks, including odometry and SLAM.

Improvements for AI systems

Here are specific improvements for AI systems based on the FoMo dataset, categorized by capability:

) 1. Robust State Estimation and Localization Under Extreme Environmental Stress

The primary improvement lies in developing localization algorithms that explicitly model and adapt to significant, non-linear environmental changes induced by seasonal variations (snow accumulation, vegetation growth).

Improvement: Integrate sensor fusion frameworks (Lidar-Inertial Odometry, Radar-Gyro Teach and Repeat) with a learned or adaptive motion model that incorporates external environmental variables. The system must be trained or fine-tuned on multi-modal data exhibiting high dynamic range in sensor readings (e.g., distinguishing between snow accumulation and terrain features).
Improved AI System Capability: An autonomous vehicle system capable of maintaining centimeter-level localization accuracy during severe winter conditions (e.g., 1m+ snow accumulation) or navigating complex off-trail terrain where GNSS signal is blocked by dense boreal forest canopy. This includes developing algorithms resilient to sensor occlusion caused by snow and vegetation.

) 2. Multi-Modal Perception for Semantic Terrain Understanding

The unique combination of sensors (Lidar, Radar, Stereo Camera, Monocular Camera) allows for richer environmental representation than single-modality systems.

Improvement: Implement deep learning models (e.g., Graph Neural Networks or Transformer architectures) that fuse sparse point cloud data from Lidar/Radar with dense visual features from stereo/monocular cameras to generate a semantically rich, temporally consistent 3D map representation (Rdm). This representation should be dynamically updated to reflect seasonal changes in surface type (pavement vs. snow vs. mud).
Improved AI System Capability: A robot capable of performing real-time terrain classification and traversability estimation. For example, the system can differentiate between a stable gravel road, deep snow banks (like those seen in the FoMo data), and steep off-trail obstacles, allowing for dynamic path planning that adjusts speed and trajectory based on predicted ground conditions.

) 3. Enhanced Sensor Calibration and Noise Modeling

The paper emphasizes extensive calibration (Lidar-to-Lidar, IMU-to-Camera) and detailed noise characterization (Allan variance analysis).

Improvement: Develop self-calibrating or online calibration modules within the localization pipeline that can estimate sensor biases (IMUs) and extrinsic transformations between disparate sensors (e.g., Radar and Lidar) in real-time, accounting for temperature-induced drift observed across the 37°C range in the dataset.
Improved AI System Capability: A navigation system with superior long-term stability. This reduces accumulated odometry drift during prolonged autonomous missions by continuously correcting for sensor degradation or environmental shifts (like temperature changes affecting sensor performance), leading to more reliable path execution over multi-day or multi-season deployments.

) 4. Robust Loop Closure and Re-localization Under Visual Degradation

The results show that Visual SLAM performs well, particularly with loop closure, but fails under specific conditions (night, heavy rain).

Improvement: Train visual place recognition models (e.g., using SuperGlue features) specifically on the multi-seasonal variations present in FoMo. The system should employ a multi-hypothesis re-localization strategy that compares feature matches across different seasonal states to prevent catastrophic failure when traditional features are obscured by snow or low light.
Improved AI System Capability: A robot capable of robust long-term autonomy in complex, changing environments where visual landmarks might be temporarily obscured (e.g., driving through dense forest during a snow event followed by clear conditions), ensuring it can successfully re-localize itself within its previously mapped area without complete failure.

) 5. Data-Driven Operational Planning and Task Adaptation

The dataset includes power consumption and motor command velocities, linking physical movement to environmental impact (e.g., immobilization).

Improvement: Develop Reinforcement Learning (RL) agents that treat the environmental state (derived from sensor fusion) as part of their observation space. The agent should learn policies for task execution that explicitly incorporate safety constraints derived from the snow data and power monitoring metadata, allowing it to dynamically choose between on-road vs. off-pavement routes based on real-time conditions.
Improved AI System Capability: An intelligent autonomous system that can perform adaptive mission planning. For instance, if the RL agent detects excessive snow accumulation or battery drain (as indicated by metadata), it can autonomously replan a path to an alternative, safer route or initiate a recovery maneuver, moving beyond simple reactive control into proactive environmental risk management.

Abstract

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.