💻 computer science

When Obstacles Bend: Modeling Vegetation Deformation in the context of Field Robotics

This paper proposes a novel approach for field robotics that characterizes vegetation through its intrinsic mechanical properties, derived from combining deformation and contact force measurements, to enable platform-independent navigation and interaction assessment.

Original authors: Muhammad Hsaeeb Zaar Khizar, Tom Montagnon, Roland Lenain, Romuald Aufrère, Johann Laconte

Published 2026-08-27

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Muhammad Hsaeeb Zaar Khizar, Tom Montagnon, Roland Lenain, Romuald Aufrère, Johann Laconte

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

In the wild, the ground is rarely a flat, empty stage. It is a tangled world of stems, leaves, and branches that do not simply block a path but yield, bend, and sometimes break under pressure. For a robot designed to move through fields, forests, or gardens, the challenge is not just seeing where an obstacle is, but understanding how that obstacle will behave when touched. A rigid wall must be avoided, but a flexible plant might be pushed aside, allowing the machine to pass without damage to itself or the crop. This distinction is critical for tasks like harvesting delicate fruit or monitoring fragile ecosystems, where the difference between success and failure lies in knowing whether a stem will snap or simply sway. To navigate this, a robot needs more than a map of shapes; it needs a sense of the material world, a way to measure the hidden strength and flexibility of the vegetation it encounters.

Researchers in France have developed a new way to give robots this sense of touch, moving beyond simple visual guesses to a physical understanding of how plants bend. Instead of asking a robot to learn how a specific machine reacts to a specific bush, the team created a method to describe the plant itself, independent of the robot pushing it. They treated the vegetation not as a static object, but as a continuous, flexible structure, similar to a long, thin rod. By combining measurements of how much force is needed to push a stem with observations of how that stem curves, they could calculate the plant's inherent stiffness. This approach allows the robot to predict how a stem will respond to a push at any height, rather than just memorizing the result of one specific collision.

The team tested their idea on two very different specimens: a small, artificial grass bush designed to mimic a real tuft of grass, and a woody twig collected from the wild. They used a robot arm equipped with a simple wire sensor that could feel the force of the push, along with a camera to watch the stem bend. As the robot slowly advanced, it pressed against the vegetation at various heights. The camera recorded the exact shape the stem took as it curved, while the sensor measured the force required to hold it in that position. From these two streams of data, the researchers reconstructed the plant's mechanical profile. They found that a stem is not uniformly stiff; it is stronger at the base and becomes more flexible toward the tip. By fitting a mathematical model to the observed bending, they could determine the plant's true stiffness along its entire length, a property that exists regardless of which robot is doing the pushing.

This method produced two different ways of describing the plant, depending on what the robot needs to do. The first, more detailed approach creates a full map of the stem's stiffness from bottom to top. This is useful for complex tasks, like assessing whether a crop has been damaged by wind or planning a harvest where the robot must interact with plants at different heights. The second approach simplifies the stem into a single number representing its overall resistance at a specific point. This "lumped" stiffness is much easier to calculate in real-time and requires only the force sensor, making it ideal for a robot that needs to make quick decisions while driving through a field at a fixed height. The researchers showed that while the simple number changes depending on where the robot touches the plant, the detailed stiffness map remains constant, offering a reliable description of the plant itself.

The experiments revealed that the simple, single-number model works well only when the robot interacts with the vegetation in a very consistent way. If the robot pushes the plant at the same height every time, the single number accurately predicts how much force will be needed. However, if the robot changes its height or the angle of approach, that single number becomes unreliable. In contrast, the detailed stiffness map, which accounts for how the stem tapers and bends along its length, remained accurate across all the different heights tested. For the woody twig, the researchers calculated a base stiffness of 2.63 newton-meters squared, while the artificial grass had a much lower stiffness of 0.112 newton-meters squared. The model successfully predicted how the stems would bend and how much force they would exert, even for heights the robot had not touched during the initial measurement.

This work suggests a shift in how robots understand their environment. Rather than learning a list of "difficult" or "easy" paths based on how a specific robot performed in the past, the robot can now learn the physical properties of the world itself. Once the stiffness of a plant is identified, that information can be transferred to any other robot, or used by the same robot with different sensors or control strategies. The researchers noted that while their current system requires a camera to see the bending, future work could link these physical properties to visual cues alone, allowing robots to "see" the stiffness of a plant without ever touching it. For now, the study provides a solid foundation for robots to move through the natural world with a genuine understanding of the materials they encounter, ensuring they can navigate, harvest, and monitor without causing unnecessary harm.

Technical Summary: When Obstacles Bend: Modeling Vegetation Deformation in the context of Field Robotics

Problem Statement

Autonomous robots operating in natural environments often must interact with vegetation rather than simply avoiding it. Current approaches to traversability typically define the environment from the robot's perspective, measuring how a specific platform responds to obstacles. This entangles the assessment of the environment with the robot's own dynamics, making the resulting characterization difficult to transfer across different platforms. Furthermore, these methods do not directly reflect the intrinsic mechanical properties of the vegetation itself, which are the true source of interaction and potential damage in applications like agriculture and environmental monitoring. Conventional representations capture the geometry of obstacles but fail to describe their mechanical behavior (e.g., distinguishing a stiff stalk from a compliant one). The paper identifies a gap in recovering a physically grounded, robot-independent mechanical description of vegetation directly from ordinary robot interactions.

Methodology

The paper proposes a framework grounded in rod theory (specifically Kirchhoff rod theory) to model vegetation stems as one-dimensional elastic continua. The approach decouples the vegetation's mechanical properties from the robot's dynamics by combining deformation measurements with contact force data. The methodology is divided into two complementary modeling branches derived from the same underlying rod theory:

Distributed Flexural-Rigidity Profile ( $EI(s)$ ):
- Model: The stem is modeled as a planar Kirchhoff rod with assumptions of planarity, quasi-static loading, inextensibility, and linear elasticity. The stem's resistance to bending is defined by its flexural rigidity, $EI(s)$ , which varies along the arc length $s$ .
- Sensing: Requires an RGB-D camera to extract the stem's centerline and curvature $\kappa(s)$ , combined with a force sensor measuring the contact force $F_c$ .
- Estimation: The internal bending moment $M(s)$ is calculated from the contact force and observed shape. The rigidity profile is estimated using the constitutive law $M(s) = EI(s)\kappa(s)$ . To address noise near the base (where curvature is low) and parameter identifiability issues, the authors parametrize $EI(s)$ using a power-law taper model: $EI(s) = EI_0 (1 - s/L)^{4p}$ . The parameters (base rigidity $EI_0$ and taper exponent $p$ ) are fitted by pooling data from contacts at multiple heights.
- Output: A spatially distributed stiffness profile that is intrinsic to the plant.
Lumped Rotational Stiffness ( $k_\theta$ ):
- Model: The stem is reduced to a rigid link attached to a compliant base via a rotational spring. This simplifies the distributed compliance into a single scalar stiffness.
- Sensing: Requires only a force sensor (specifically a wire-based sensor) and robot pose data; no vision pipeline is needed.
- Estimation: The contact force $F_c$ and contact height $h$ generate a torque $\tau = F_c h$ . The deflection angle $\theta$ is recovered geometrically from the robot's pose and the wire sensor's elongation. The stiffness is derived from the torque balance $\tau = k_\theta \theta$ .
- Output: A single scalar stiffness value representing the aggregate resistance at a specific contact height.

Key Contributions

The paper presents three main contributions:

A Unified Rod-Theoretic Framework: A single theoretical foundation that captures both smooth bending along the stem's length and near-rigid rotation about a compliant base, allowing for the estimation of either a distributed profile or a lumped parameter.
A State Estimation Pipeline: A method to recover underlying mechanical parameters directly from interaction data (force and shape) without requiring offline calibration, simulation, or destructive testing.
Experimental Evaluation: A comparative analysis of the two proposed models, highlighting their respective strengths, limitations, and applicability to different robotic tasks.

Experimental Results

The framework was validated on two specimens: an artificial grass bush and a woody twig, pushed by a UR10 arm equipped with a wire-based force sensor and an RGB-D camera.

Flexural-Rigidity Profile ( $EI(s)$ ):
- The direct estimator (Equation 7) suffered from high noise near the base where curvature approaches zero.
- The parametric taper-law model successfully smoothed these estimates. For the twig, the fitted parameters were $p=0.43$ and $EI_0=2.63$ N·m²; for the grass, $p=1.60$ and $EI_0=0.112$ N·m².
- Generalization: The estimated $EI(s)$ profile remained stable and consistent across different contact heights. When used in forward simulations, the model accurately predicted forces and shapes for pushes at heights not included in the fitting process (median shape RMSE of 14.3 mm for the twig and 3.9 mm for the grass).
Lumped Stiffness ( $k_\theta$ ):
- The model successfully estimated stiffness using only force data.
- Limitation: The estimated $k_\theta$ varied significantly with contact height (e.g., decreasing from 1.60 to 0.98 N·m/rad for the twig as height increased).
- Behavior: The linear torque-balance model could not capture the non-linear yielding behavior observed when vegetation bends sufficiently to allow traversal (force plateau).

Significance and Claims

The paper claims that characterizing vegetation through its intrinsic mechanical properties, rather than platform-dependent metrics, allows for vegetation-aware navigation that is transferable across different robots, sensors, and control policies.

Transferability: The distributed flexural-rigidity profile $EI(s)$ is presented as a robot-independent descriptor. Once identified, it can be reused across different interaction geometries and robots without re-estimation.
Task Suitability:
- The lumped model is suitable for real-time traversability decisions where the robot interacts at a fixed, known height and within a limited elastic range, as it requires only onboard force and pose data.
- The distributed model is necessary for tasks requiring detailed mechanical understanding (e.g., crop-lodging assessment, forestry understory characterization) or where interaction geometry varies (e.g., harvesting), as it captures the full spatial complexity of the stem.
Modesty: The authors acknowledge that the lumped model is insufficient for fully characterizing vegetation response across all configurations due to its dependence on interaction geometry and inability to model non-linear yielding. They also note that while the method applies to any deformable entity, the current validation is limited to specific specimens and does not yet include a dataset linking visual observations to mechanical properties for learning-based prediction.