Multi-Reference Path Tracking Control for an Agricultural Tractor with Nonlinear Model Predictive Control

arXiv:2610.00057 · cs.RO · Submitted 2026-09-04 · Read on arXiv

Listen

Radio episode about this paper

Transcript

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

Rosa: Today's paper: "Multi-Reference Path Tracking Control for an Agricultural Tractor with Nonlinear Model Predictive Control".

Dev: Guiding agricultural tractors along predefined paths is crucial for precision agriculture,

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

Title and authors: Rosa: So we're looking at this paper titled "Multi-Reference Path Tracking Control for an Agricultural Tractor with Nonlinear Model Predictive Control," and the authors are Marcel Moll and Timo Oksanen from the Technical University of Munich. I think the title immediately tells us they're tackling a complex problem in precision agriculture by using nonlinear model predictive control to guide tractors along paths.

Dev: Yeah, it looks like they’ve taken an existing path tracking method and significantly upgraded it by adding this multi-reference approach directly into the cost function of their MPC formulation. That suggests they are aiming for better handling of the geometry than what simpler models can manage on curved terrain.

Taro: From an autonomy standpoint, having a method that explicitly manages multiple reference segments within the optimization objective is interesting because real-world paths aren't always perfectly smooth or linear; this hints at a need for more adaptive control when the environment deviates from the ideal plan.

Rosa: Exactly, and I’m wondering if they tested this on actual fields outside of a controlled lab setting, because that’s where you really see if these complex algorithms hold up against dirt and varying conditions.

Dev: The abstract mentions field testing with a tractor controlled via the Tractor Implement Management steering interface, which gives us some real-world validation data to look at when we talk about performance under operational stress.

Taro: If this approach can handle the transition between segments smoothly, it opens up possibilities for autonomous machinery to navigate complex boundaries or uneven terrain where traditional single-reference trackers would struggle with those sharp changes in direction.

Rosa: It sounds like they are focusing on making sure the tractor doesn't just follow one line perfectly, but intelligently manages switching between several lines when necessary.

Dev: That transition management is key; if the system has to make a sudden switch between references, we need to know how quickly that happens and what kind of stability we can expect in the control loop rate.

The paper's summary: Rosa: To summarize what they did, this paper develops an advanced path tracking controller using Nonlinear Model Predictive Control where they incorporate multiple segments of a piecewise-linear reference path right into the objective function to improve how the tractor follows those paths.

Dev: That means instead of just penalizing the distance from one single line, the cost function now considers several potential lines, which should help it manage curves much more effectively than previous methods.

Taro: I see how that addresses the issue mentioned in earlier literature where single-reference controllers suffer from those saw-tooth tracking errors when paths are curved; this multi-segment inclusion is specifically designed to give the controller a better predictive capability.

Rosa: Right, and they introduced a specific mathematical way to approximate those changes between segments using sigmoid functions, which they claim makes the reference function non-smooth at the segment boundaries manageable for Newton-type optimizers.

Dev: The formulation uses a cost term that is essentially a nonlinear least-squares expression involving both the current state and control inputs, which is typical for MPC but adapted here to incorporate this multi-reference distance metric.

Taro: The way they define progress s based on the segment endpoints, using the formula shown in equation (five), seems important because it ensures that we only consider a point on the path when it actually lies within the bounds of a specific line segment.

Rosa: So, essentially, they’ve built a system where you can feed it several lines simultaneously and use these sigmoid functions to blend between them smoothly during the optimization process.

Dev: That blending mechanism is what I'm most interested in from an engineering standpoint; we need to ensure that the parameter k controlling the steepness of that transition doesn't introduce oscillations or instability into the solver when things get tight.

The paper's improvements: Rosa: The authors suggest several key improvements, primarily focusing on how they handle those multiple segments and how they select which reference segment to follow at any given moment during the entire path.

Dev: They introduce a path handler that does preprocessing steps, like making sure all segments are long enough and generating a starting trajectory, which is necessary because you can't just start tracking randomly on a complex path.

Taro: The selection mechanism is interesting; they calculate progress s in real time using the formula involving coordinates to determine which two adjacent segments are relevant when updating the path index, and this prevents skipping over regions where the path geometry is changing rapidly.

Rosa: So, it’s not just about tracking multiple lines; it's also about having a smart system that dynamically figures out which part of the path is most accurate for the current tractor position.

Dev: They also include a cost term that penalizes excessive control effort, specifically the steering-angle derivative, which lets them tune how much they prioritize smooth motion versus aggressively correcting tracking errors.

Taro: That control effort penalty is vital because it balances precision with physical limitations; if you don't include that term, the controller might try to correct every tiny error too hard and end up with jerky movements.

Rosa: It seems like the authors are presenting a comprehensive solution that combines a robust cost function for tracking, smart segment selection logic, and an effort-aware steering rate penalty all in one system.

Dev: That integrated approach makes the controller much more adaptable to different path topologies than just having a static set of rules; it lets the system react to what’s actually happening on the ground.

Conclusion: Rosa: So, wrapping up on this paper, we see that by incorporating multiple segments of a piecewise-linear reference path into the objective function via sigmoid functions, this Multi-Reference Path Tracking Control for an Agricultural Tractor with Nonlinear Model Predictive Control provides a way to track curved paths much more reliably.

Dev: And the real-time selection of viable reference segments based on progress s seems to be the mechanism that makes this work practically viable outside of simulation.

Taro: I think the implication here is that for autonomous systems, especially in agriculture, we can expect a level of path following accuracy on complex geometries that was previously thought to require much more sophisticated and computationally expensive planning methods.

Rosa: And while they report a mean absolute cross-track error of six point one cm during their field test on a figure-eight shaped path, which sounds quite good for an outdoor application, we still need to see how long this system can maintain that performance over extended operational hours.

Dev: That three point four five ms convergence time is decent for a real-time loop, but the ultimate success hinges on how well the system handles actuator delays and latency in a real tractor environment when it’s running at those operational frequencies.

Taro: If this method proves robust when the world misbehaves—say, if a crop row shifts unexpectedly—it sets a new baseline for how resilient agricultural machinery can be to dynamic path changes.

Technical University of Munich

cs.RO

Submitted: 2026-09-04

Updated: 2026-09-04

License: http://creativecommons.org/licenses/by-nc-nd/4.0/

Importance score: 77/100

The gist: Guiding agricultural tractors along predefined paths is crucial for precision agriculture, and this study develops an advanced path tracking controller using Nonlinear Model Predictive Control (NMPC)

Key concepts

Nonlinear Model Predictive Control (NMPC)
NMPC is an optimization method that predicts future system behavior over a time horizon and calculates the best control inputs to minimize a cost function. It handles complex, nonlinear systems by solving an optimization problem repeatedly in real-time, allowing for proactive path correction.
Multi-Segment Reference Cost Function
This technique involves including multiple pieces of the desired path into the controller's cost calculation using sigmoid functions. This forces the controller to focus on tracking only the segment closest to its predicted position, which is crucial for accurately following curves.
Piecewise-Linear Reference Path
Instead of a single straight line, this path is divided into several short, straight segments. The controller uses these segments as targets. By incorporating multiple segments into the cost function, the system gains better predictive capability when navigating complex or curved routes.
Path Handler
This component manages all the different reference path segments for the tractor. It performs preparatory tasks like checking segment lengths and generating a starting trajectory (like a straight line or Dubins path). It also updates which two adjacent segments are currently relevant for tracking, ensuring smooth transitions.

Terminology

Summary

Guiding agricultural tractors along predefined paths is crucial for precision agriculture, and this study develops an advanced path tracking controller using Nonlinear Model Predictive Control (NMPC) that incorporates multiple segments of a piecewise-linear reference path directly into its objective function. The core contribution lies in extending existing MPC methods to handle curved paths effectively by utilizing multi-reference predictions, thereby improving the controller's predictive capability and tracking performance.

The gist

This study develops a path tracking controller based on Nonlinear Model Predictive Control, which incorporates multiple segments of a piecewise-linear reference path directly into the objective function.

Nonlinear Model Predictive Control Formulation

The MPC problem is formulated as:

min

x(·),u(·)

Z tf

0

L x(t),u(t), p(t)

dt + E (x(tf), p(tf)) subject to x˙(t) = f(x,u, p), x˙=f. The cost function to be minimized consists of a Mayer term E and a Lagrange term L, where the latter is formulated as a nonlinear least-squares expression:

L x(t),u(t), p(t) = 1/2 (y x(t),u(t), p)(yref - yx(t),u(t), p) T W.

System Model and Control Inputs

The underlying system model utilized is a continuous time kinematic bicycle model, which includes an additional state for the steering angle derivative ˙δ t. The nonlinear system dynamics are given by:

f(x,u, p) = [x˙= [v t cos(ψ g), v t sin(ψ g), v t tan(δ t), lw ˙δ t], u = ˙δ t]. The control variable u used by the controller is the steering-angle derivative. For this study, the velocity v t is assumed to be manually controlled by the operator and treated as a constant parameter that may change between controller iterations.

Multi-Segment Reference Cost Function

To enhance predictive capability on curved paths, multiple segments of a piecewise linear reference are included into the cost function using sigmoid functions σ(s, s0) = 1 / (1 + exp−k (s − s0)). The combined distance cost is then defined as:

cd˜ = 1 − σ(s1, 1)˜d1 + nsegX−1 i=2 h βiαi 1 − σ(si, 1)˜di + βnseg αnseg σ(snseg, 0)˜dnseg. This formulation ensures that the desired behavior of the distance cost term cd˜ is to yield the cross-track metric for exactly one reference segment, namely the one closest to the predicted position.

Path Handler and Runtime Parameter Updating

A path handler is implemented to manage multiple reference segments within a predefined path. This handler performs preprocessing steps such as ensuring all segments exceed a minimum length and generating a startup trajectory (either straight connection or Dubins path). In every controller iteration, the path handler maintains the current path index by calculating the progress s using: s = [(x − x1) (x2 − x1) + (y − y1) (y2 − y1)] / [(y2 − y1) squared + (x2 − x1) 2]. This ensures that only two adjacent segments are considered when updating the path index, preventing the path from being skipped in overlapping regions.

Real-time Implementation Details

The NMPC framework is implemented using acados, which offers a convenient interface for formulating optimal control problems with CasADi. The QP solver selected is HPIPM (Frison and Diehl (2020)) to efficiently handle the proposed cost formulation. The prediction horizon is set to 100 steps with a fixed interval of 100 ms, corresponding to the maximum control frequency over an ISO 11783 CAN bus. To mitigate actuator delays tdδ, the system dynamics are propagated using a Runge–Kutta integrator, applying previous control commands according to the estimated delay while assuming constant tractor speed during prediction. The controller is evaluated on a Fendt 314 Vario Gen 4 tractor equipped with a TIM interface and GNSS antennas providing RTK fixes. The NMPC solver converged on average after 3.45 ms and tracked the curved reference path with a mean absolute cross-track error of 6.1 cm during field testing.

Results Summary

The controller was tested on a figure-eight shaped path, achieving an average cross-track error of 0.2 cm with a standard deviation of 8.4 cm, and an absolute value of the cross-track metric evaluated to 6.1 cm.

Improvements for AI systems

Here are specific improvements that can be made to AI systems, based on the methodology presented in this paper, along with what those improved systems could achieve:

The core improvement lies in upgrading existing path-following or autonomous vehicle control algorithms from single-reference tracking to robust, multi-segment path following.

  1. Development of a novel controller that integrates a Multi-Reference Model Predictive Control (MR-NMPC) framework into existing trajectory planning modules for autonomous vehicles (e.g., drones, ground robots, or agricultural machinery).

  2. Implementation of a dynamic reference segment selection mechanism based on real-time vehicle progress along the path, utilizing sigmoid transition functions to ensure smooth switching between reference lines without causing non-smooth optimization issues.

  3. Incorporation of a combined cost function that simultaneously minimizes cross-track error against the active segment and penalizes excessive control effort (steering rate derivative), allowing for a tunable trade-off between ride smoothness and aggressive error correction.

This improved AI system can perform the following specific tasks:

  1. Automated guidance along complex, predefined routes in unstructured environments (e.g., navigating field boundaries or crop rows).

  2. Accurate tracking of piecewise linear paths that change direction frequently, such as those found in obstacle avoidance maneuvers or precision agricultural boundary following.

  3. Robust operation on curved paths where traditional single-reference controllers suffer from saw-tooth tracking errors, enabling smoother and more precise maneuvering around obstacles or along curvilinear infrastructure.

  4. Real-time adaptation to path topology changes by dynamically selecting the most appropriate reference segment as the vehicle progresses, ensuring continuous and optimized trajectory adherence even when the reference geometry is complex.

Abstract

Guiding a tractor along a predefined reference path is a key component of precision agriculture. This study develops a path tracking controller based on Nonlinear Model Predictive Control, which incorporates multiple segments of a piecewise-linear reference path directly into the objective function. In addition, methods for selecting viable reference segments from the full path are presented. The control system is evaluated during a field test with a tractor controlled via the Tractor Implement Management steering interface. The NMPC solver converged on average after 3.45 ms and tracked the curved reference path with a mean absolute cross-track error of 6.1 cm.

Related papers