⚡ electrical engineering

Koopman Operator Framework for Modeling and Control of Off-Road Vehicle on Deformable Terrain

This paper presents a hybrid physics-informed and data-driven framework that utilizes the Koopman operator to transform complex, computationally intensive terramechanics models into a linear system, enabling robust and stable model predictive control for autonomous off-road vehicles operating on deformable terrain.

Original authors: Kartik Loya, Phanindra Tallapragada

Published 2026-04-01

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Kartik Loya, Phanindra Tallapragada

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

Imagine you are trying to teach a robot car how to drive through a muddy field. On a paved road, driving is easy: you turn the wheel, and the car turns. The tires grip the asphalt, and the physics are predictable.

But off-road? It's a different beast. When a tire hits soft mud or sand, the ground squishes, the tire sinks, and the mud pushes back. The car doesn't just turn; it slides, sinks, and fights the terrain. This is called deformable terrain.

The problem is that the math required to predict exactly how a tire will sink into mud is incredibly complex. It's like trying to calculate the exact movement of every single grain of sand in a hourglass while the car is driving. If you try to use this super-detailed math to control the car in real-time, the computer gets overwhelmed and the car crashes.

This paper presents a clever solution: The "Magic Crystal Ball" (The Koopman Operator).

Here is the breakdown of their approach using simple analogies:

1. The Problem: The "Over-Engineered" Map

Traditional engineers try to build a perfect map of the mud. They measure the soil's stickiness, the tire's weight, and the speed. This is the Bekker-Wong-Reece model. It's accurate, but it's so heavy and slow that a computer can't use it to steer the car while it's moving. It's like trying to navigate a city by calculating the friction of every single brick on the sidewalk before you take a step.

2. The Solution: The "Linear Shortcut"

The authors used a mathematical trick called the Koopman Operator.

The Analogy: Imagine you are watching a chaotic dance floor. It's hard to predict where everyone will be next because people are bumping into each other and spinning.
The Trick: Instead of tracking every person's chaotic movement, you lift the camera up to the ceiling and look at the pattern of the crowd as a whole. From this new angle, the chaos looks like a simple, straight line.
In the Paper: They took the messy, non-linear physics of the mud and "lifted" it into a higher-dimensional space where the rules become simple and linear (like a straight line). This allows them to use fast, simple math to predict where the car will go next.

3. Learning from "Video Games" (Simulation)

Since they can't drive a real car into a mud pit 1,600 times to collect data (that would be expensive and dangerous), they built a virtual simulator.

They created a "digital twin" of the car and the mud.
They ran thousands of virtual drives on Sandy Loam (grainy sand) and Clay (sticky mud).
They fed all this data into their "Magic Crystal Ball" (the Koopman model) to teach it how the car behaves on each specific soil type.

4. The "Smart Filter" (Grassmannian Distance)

One of the coolest parts of their method is how they handle new data.

The Analogy: Imagine you are teaching a student. If you show them a picture of a cat, then another picture of a cat, then another, they aren't learning anything new. But if you show them a picture of a dog, that's new information.
The Tech: The system uses a "Grassmannian distance" test. It's like a smart filter that asks: "Is this new driving data actually teaching me something new, or is it just a repeat of what I already know?" If it's a repeat, the system ignores it to save time. If it's new, it updates the model. This makes the system efficient and able to learn continuously without getting bogged down.

5. The Result: Driving the "Aggressive Maneuvers"

They tested this system by putting it inside a Model Predictive Controller (MPC). Think of the MPC as a super-smart co-pilot who looks 2 seconds into the future.

The Test: They asked the car to do a "Fishhook" maneuver (a sharp turn followed by an even sharper turn back) on soft soil.
The Outcome:
- On Sandy Loam: The car followed the path perfectly, adjusting its speed and steering instantly.
- On Clay: The car also did great, but only because they used the Clay-specific model.
- The Warning: When they tried to use the "Sandy Loam" model to drive on "Clay," the car crashed. It thought the mud was slippery sand, but it was actually sticky clay. The car spun out.

The Big Takeaway

This paper proves that you don't need a super-computer to drive off-road. You can use a "smart shortcut" (the Koopman operator) that learns from simulations and updates itself with new data.

However, there is a catch: You need a specific "shortcut" for every type of mud. You can't use the map for sand to drive on clay. The system needs to know exactly what kind of ground it is on to predict the future correctly.

In summary: They turned a messy, impossible-to-solve mud problem into a clean, fast, linear prediction problem, allowing autonomous vehicles to drive aggressively and safely through soft terrain without getting stuck or crashing.

1. Problem Statement

Autonomous off-road vehicle control faces significant challenges distinct from on-road autonomy due to the complex physics of deformable terrain.

The Core Issue: On soft soil, tire-terrain interaction involves non-linear phenomena such as soil compaction, shear deformation, wheel sinkage, and slip. These factors couple traction and steering to soil strength in ways that standard rigid-surface models cannot capture.
The Trade-off: High-fidelity terramechanics models (e.g., Discrete Element Methods or detailed Bekker-Wong-Reece formulations) are accurate but computationally expensive, making them unsuitable for real-time Model Predictive Control (MPC). Conversely, simplified rigid-body models fail to predict vehicle behavior on soft soil, leading to tracking errors, excessive slip, or loss of controllability.
The Goal: Develop a computationally efficient, linear, and updateable model that captures the dominant non-linear wheel-terrain dynamics to enable real-time predictive control on deformable terrain.

2. Methodology

The authors propose a hybrid physics-informed and data-driven framework utilizing the Koopman Operator Theory.

A. High-Fidelity Simulation & Data Generation

Vehicle Model: A 5-Degree-of-Freedom (5-DOF) vehicle model combining a single-track (bicycle) dynamic model for planar motion and a half-car suspension model for vertical/pitch dynamics.
Terramechanics: The wheel-soil interaction is modeled using the Bekker–Wong–Reece theory. This includes:
- Pressure-sinkage relations (Bekker) to determine normal forces and sinkage.
- Shear stress relations (Wong-Reece) to calculate longitudinal and lateral forces based on soil cohesion, friction, and shear modulus.
Dataset: Large-scale simulations were run on sandy loam and clay soils. 1,600 diverse trajectories (including slaloms, fishhooks, and circles) were generated for each soil type using randomized inputs to ensure persistence of excitation.

B. Koopman Operator Identification

The non-linear system dynamics are lifted into a higher-dimensional space where they evolve linearly.

Framework: The system is approximated as a linear time-invariant (LTI) system in a "lifted" state space ( $z$ ):
$z_{t+1} = A z_t + B u_t$
$y_t = C z_t$
Recursive Subspace Identification (SSID): Instead of batch processing massive datasets (which is memory-intensive), the authors use a recursive subspace identification method.
- Incremental Learning: The algorithm updates the observability subspace as new data segments arrive, avoiding the need to re-factorize large Hankel matrices.
- Grassmannian Distance: A key innovation is the use of Grassmannian distance to measure the similarity between the current data subspace and the new incoming data. Only "informative" data segments (those with a distance greater than a threshold $\epsilon$ ) are used to update the model. This prevents redundancy and ensures the model learns diverse dynamics without bloating memory.
Lifting Map: Gaussian Processes (GP) are trained to map the physical vehicle states ( $x$ ) to the latent Koopman coordinates ( $z$ ).

C. Control Strategy: Constrained Koopman-MPC

The identified linear Koopman model is embedded into a Model Predictive Control (MPC) scheme.

Prediction: The linear Koopman predictor forecasts future states over a finite horizon.
Kinematic Update: Global position ( $X, Y$ ) and heading ( $\psi$ ) are updated using kinematic integration of the predicted body-frame velocities.
Optimization: The MPC solves a constrained quadratic program to minimize tracking error and control effort while respecting actuator limits (steering angle $\delta$ and drive torque $\tau$ ).

3. Key Contributions

Hybrid Modeling Framework: Successfully bridges the gap between high-fidelity physics (Bekker-Wong-Reece) and real-time control by creating a linear surrogate model via Koopman operators.
Recursive Identification with Data Selection: Introduces a scalable recursive SSID algorithm enhanced by Grassmannian distance filtering. This allows the model to be updated incrementally with new data segments while rejecting redundant information, making it suitable for online adaptation.
Soil-Specific Operators: Demonstrates that distinct Koopman operators are required for different soil types (sandy loam vs. clay) due to the fundamental differences in shear and sinkage dynamics.
Real-Time MPC Implementation: Validates the framework in a constrained MPC setting, achieving stable tracking of aggressive maneuvers (e.g., fishhook) on deformable terrain with solver times ( $\approx 0.085$ s) well within the control sampling period ($0.1$s).

4. Results

Prediction Accuracy:
- The models achieved stable short-horizon accuracy. For sandy loam, the model order was 18; for clay, it was 19.
- Robustness: Models trained on flat terrain maintained high accuracy on mildly uneven terrain (up to 0.1m elevation variation) without retraining.
- Error Growth: Lateral velocity ( $v$ ) showed the fastest error growth in open-loop prediction, while longitudinal velocity ( $u$ ) and yaw rate ( $\dot{\psi}$ ) remained more stable.
Soil Specificity (Crucial Finding):
- Using a Koopman operator trained on sandy loam to control a vehicle on clay resulted in a massive performance degradation (RMSE increased by ~90-150%).
- In MPC tests, the "wrong" soil model caused the vehicle to drift off the reference trajectory and saturate steering actuators, leading to a 3x increase in the cost function compared to the correct soil-specific model.
Maneuver Diversity: The models generalized well across various maneuvers (straight, circle, slalom, fishhook) within the same soil type, indicating the dataset provided sufficient coverage of the state space.

5. Significance and Future Directions

Significance: This work provides a viable pathway for real-time autonomous control on soft soil, a domain previously dominated by either slow, high-fidelity simulations or inaccurate, simplified models. The recursive nature of the identification allows for potential online adaptation as a vehicle encounters new terrain conditions.
Future Work:
- Extending the framework to handle larger terrain elevation variations and mixed soil conditions.
- Developing ensemble methods where multiple Koopman operators (for different soils) are blended or selected online based on real-time terrain estimation.
- Integrating physics-informed constraints directly into the Koopman operator learning process to further improve generalization and data efficiency.

In conclusion, the paper demonstrates that Koopman operator theory, when combined with recursive subspace identification and physics-based simulation, offers a powerful, computationally efficient solution for the challenging problem of off-road vehicle control on deformable terrain.