💻 computer science

Tracking the Ground: Online Lidar Identification of Robot-Induced Soil Deformation in Agricultural Environments

This paper proposes an online framework using LiDAR observations and a reduced-order parametric model to quantify and track robot-induced soil deformation in real-time, enabling soil-aware robotic operations that can adapt behaviors to minimize agricultural soil degradation.

Original authors: Tom Montagnon, Johann Laconte, Benoit Thuilot, Wonjae Cho, Roland Lenain

Published 2026-09-16

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Tom Montagnon, Johann Laconte, Benoit Thuilot, Wonjae Cho, Roland Lenain

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The soil beneath a farmer's feet is not merely dirt; it is a living, breathing foundation that holds water, anchors roots, and feeds the world. For decades, the drive to feed a growing population has pushed agriculture toward larger, heavier machinery. While these machines are efficient, their weight and the sheer repetition of their tracks can crush the delicate structure of the earth, turning loose, healthy ground into hard, compacted slabs that choke plant life. This damage is cumulative, meaning even a single pass might seem harmless, but repeated trips over the same path can permanently degrade the land. In recent years, the agricultural world has looked to robotics as a solution, hoping that lighter, autonomous machines could work the fields without causing this harm. However, a new realization has emerged: even a lightweight robot can damage the soil if it drives over the same spot too often, especially if the ground is wet or loose. The challenge, then, is not just building a lighter machine, but teaching that machine to "feel" the ground in real time, understanding exactly how its own wheels are changing the terrain beneath it.

This is the precise problem a team of researchers from France and Japan set out to solve. They developed a system that allows a robot to watch the ground it is driving on and instantly calculate how much it is sinking or shifting the soil. Instead of relying on pre-made maps or stopping to take physical samples, the robot uses two laser scanners—one mounted at the front and one at the back. As the robot moves forward, the front scanner sees the ground before the wheels touch it, while the rear scanner sees the same patch of ground immediately after the wheels have passed. By comparing these two views, the robot can see the exact difference the wheels made, measuring the tiny dips and mounds created by the passage.

The researchers did not stop at simply measuring the change; they built a mathematical model that explains why the ground changed. They reduced the complex physics of soil behavior down to three simple, observable numbers. The first number describes how deep the wheel sinks into the earth. The second measures how much of the soil is squeezed tightly together versus how much is pushed aside to the side of the track. The third describes how the loose piles of displaced soil naturally settle and slide until they reach a stable slope. By continuously updating these three numbers as the robot drives, the system creates a live, evolving picture of the soil's condition. It is like the robot is constantly asking, "How soft is the ground right here, and how is it reacting to my weight?"

To test this idea, the team drove their robot across two very different types of terrain. In one test, they drove over a field of freshly watered soil, which was soft and prone to deep ruts. In the other, they drove over a gravel surface that had been packed down by previous traffic. In the wet field, the system correctly identified that the soil was soft, showing deep sinking and very little soil being pushed to the sides. As the robot moved through the wettest part of the field, the model showed the sinking depth increasing and then decreasing, matching the pattern of the irrigation. On the gravel, the system detected something different: the soil was not just sinking, but the loose stones were being pushed aside and piling up, a behavior the model captured by adjusting its "pushed aside" parameter to a value greater than one, indicating the ground was loosening as it was moved.

The results showed that the robot's model could predict the shape of the ruts with remarkable accuracy, matching the actual measurements taken by the sensors almost perfectly. The system was able to distinguish between the two very different soil types and adapt its understanding of the ground as it moved from one condition to another. Crucially, the researchers found that the model worked well even when they simplified it, suggesting that the core three parameters are enough to describe the soil's behavior for the purpose of navigation. The system operates fast enough to keep up with the robot's movement, processing the data in less than a second for each section of the path.

This work represents a significant step toward truly "soil-aware" robotics. By giving the machine the ability to quantify the damage it is causing in real time, the technology lays the groundwork for future robots that can automatically adjust their behavior to protect the land. A robot equipped with this system could theoretically slow down, change its path, or alter its weight distribution the moment it senses the ground is becoming too soft or too damaged. While the current study focuses on measuring and understanding the deformation, the ultimate goal is to use this knowledge to prevent the damage before it happens, ensuring that the machines of the future work in harmony with the living soil they depend on.

Technical Summary: Tracking the Ground – Online Lidar Identification of Robot-Induced Soil Deformation

Problem Statement
Agricultural robotics aims to improve field efficiency and sustainability, yet vehicle-soil interactions remain a critical challenge. Repeated traffic, even by lightweight autonomous robots, induces soil compaction, rutting, and structural degradation, which compromises root development and long-term productivity. While existing approaches like Controlled Traffic Farming limit damage by confining wheels to permanent lanes, they do not address the cumulative damage caused by trajectory repetition on variable soil. Furthermore, current soil-aware planning methods often rely on static susceptibility maps or empirical update laws rather than direct observation of deformation. A significant gap exists in the ability to explicitly quantify soil state as a variable during operation using non-invasive sensing, specifically to enable robots to adapt their behavior based on real-time soil response.

Methodology
The authors propose a framework to quantify traffic-induced soil deformation and estimate its evolution online using dual-lidar observations. The approach consists of three main stages:

Data Acquisition and Processing:
- Dual-Lidar Setup: The system employs a front sensor ( $fL$ ) to capture the terrain before wheel contact and a rear sensor ( $bL$ ) to capture the same area after passage.
- Registration and Mapping: Raw scans are fused using offline extrinsic calibration. An Iterative Closest Point (ICP) based mapper creates a sparse global reference map ($wPref$) and trajectory ($wTref$).
- Heightmap Generation: Both front and back point clouds are independently registered to the global reference frame to eliminate drift. The data is spatially filtered to a Region of Interest (ROI) around the trajectory and rasterized into regular heightmaps ( $wHf$ and $wHb$ ).
- Displacement Calculation: The difference map ( $wH\Delta = wHb - wHf$ ) represents the ground deformation.
Reduced-Order Parametric Soil Model:
Instead of computationally expensive methods like Finite Element Method (FEM) or Discrete Element Method (DEM), the authors introduce a parametric model defined by three physically interpretable parameters:
- Sinkage Depth ( $d$ ): Represents the effective normal response of the soil (analogous to Bekker's pressure-sinkage relationship), identified directly from lidar data rather than requiring offline calibration.
- Displacement-Compaction Ratio ( $\chi$ ): The proportion of displaced soil volume that is pushed laterally versus compacted vertically. Values $>1$ indicate soil dilatancy (loosening).
- Angle of Repose ( $\beta$ ): Governs the surface relaxation of displaced soil heaps to a mechanically stable state, modeled via a discretized Exner equation.
- Simulation: The model simulates the process of wheel sinking, material redistribution, and surface relaxation to predict the post-traversal heightmap ($wHsim$).
Online Parameter Estimation:
- Sliding-Window Optimization: The trajectory is divided into short, overlapping segments. For each segment, the model parameters ( $\theta = \{d_1, d_2, \chi, \beta\}$ ) are estimated by minimizing the weighted $L_1$ cost function between the simulated heightmap ($wHsim$) and the measured rear heightmap ( $wHb$ ).
- Gradient Descent: Since the simulation lacks an analytical Jacobian, gradients are approximated via forward finite differences. Parameters are updated using gradient descent with bounds, ensuring errors do not accumulate along the trajectory.

Key Contributions

Lidar-Based Processing Pipeline: A novel method to quantify robot-induced soil deformation through differential heightmap analysis using a dual-lidar configuration.
Reduced-Order Parametric Model: A compact soil deformation model with three physically interpretable parameters ( $d, \chi, \beta$ ) that can be identified online from sequential lidar data without invasive calibration.
Experimental Validation: Demonstration of the approach's applicability across diverse soil conditions (watered agricultural soil and packed gravel), showing the ability to capture deformation and track parameter evolution.

Results
Experiments were conducted on a four-wheel, dual-steering axle robot (700 kg) traversing two distinct environments: "Watered" (irrigated seedbed) and "Gravel" (compacted surface).

Parameter Evolution: The identified parameters successfully reflected soil conditions. On watered soil, sinkage depth and angle of repose followed the irrigation profile, while $\chi$ remained near zero (indicating compaction without significant lateral displacement). On gravel, $\chi$ increased beyond 1 (indicating decompaction/dilatancy), and the angle of repose evolved with observed deposition.
Accuracy: An ablation study showed the proposed model achieved a residual cost matching the measurement noise floor ( $\approx 5.12 \times 10^{-3} m^3 m^{-2}$ ). The model's prediction cost (evaluated on unseen terrain) remained close to this noise level.
Ablation Findings: Merging left and right sinkage depths into a single parameter improved forward prediction efficiency with negligible loss in accuracy, suggesting separate depths are primarily useful for mapping observed deformation rather than predicting future states. Limiting surface relaxation iterations reduced computation time by a factor of three with a minor accuracy penalty (5–8% cost increase), suitable for real-time control.
Computational Performance: At a cell size of $\Delta x = 0.02$ m, computation times were approximately 1 second, compatible with the low speeds of agricultural operations. Coarsening the grid beyond 0.045 m offered diminishing computational returns while significantly degrading accuracy.

Significance and Claims
The paper claims to establish a foundational prerequisite for "soil-aware robotic operation." By making soil response measurable and interpretable during operation, the framework provides a continuously updated, physically grounded representation of soil state. The authors explicitly state that this work focuses on establishing the ability to identify soil parameters online; the exploitation of these estimates to adapt robotic behaviors (e.g., trajectory planning, traction control) to reduce degradation is left for future work. The proposed method bridges the gap between high-fidelity but slow terrain models and static, non-observational planning proxies, enabling robots to potentially adapt to cumulative soil damage in real-time.

Technical Summary: Tracking the Ground – Online Lidar Identification of Robot-Induced Soil Deformation

More like this