Source-linked AI summary
Multi-Reference Path Tracking Control for an Agricultural Tractor with Nonlinear Model Predictive Control
Marcel Moll, Timo Oksanen
TL;DR
A cost-based NMPC path tracker limited to one linear reference segment can misrepresent curved paths over longer prediction horizons. This paper extends the controller to multiple reference segments and tests it in a tractor field experiment, reporting a 6.1 cm absolute cross-track metric and a 3.45 ms average solution time.
Problem
A cost-based path-tracking controller limited to one linear reference segment can misrepresent curved paths, producing saw-tooth error and limiting longer prediction horizons.
Source: §1.1 · p. 2Method
The paper extends an NMPC path-tracking controller to incorporate multiple reference segments within the prediction horizon and implements it with custom path-segment handling and a TIM steering interface.
Source: §1.2 · p. 2Results
The field test’s absolute cross-track metric was 6.1 cm, and the NMPC solver returned a solution after an average of 3.45 ms.
Source: §4.1 · p. 5, §4.2 · p. 6Takeaways & Limitations
The controller tracked the curved reference path in a field experiment, and its real-time capability with a long prediction horizon seems promising.
Source: §4.1 · p. 5, §4.2 · p. 6, §5 · p. 6The evaluated setup assumes that the operator manually controls tractor velocity, which is treated as a constant parameter that may change between controller iterations.
Source: §2.2 · p. 2
Abstract
from arXiv · showhide
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.
1. INTRODUCTION
Automatic guidance systems support tractor navigation along predefined paths, while MPC provides an online optimization approach for control. This work extends path tracking to use multiple reference segments within the prediction horizon.
1. INTRODUCTION
- Centimeter-level GNSS accuracy has helped make automatic guidance systems integral to precision agriculture.The introduction describes following a predefined path with minimal deviation as the guidance system’s main task.Source: §1 · p. 1
1.1 Current State of Research
- MPC computes control outputs through online optimization over a limited prediction horizon, while path tracking avoids forcing a vehicle to catch up to a timed schedule.The paper distinguishes space-based path tracking from time-based trajectory tracking as more suitable when exact timing is secondary.Source: §1.1 · p. 1
- A cost-based controller limited to one linear reference segment can produce saw-tooth tracking error on curves and cannot correctly represent the path over longer horizons.The cited formulation leaves the kinematic system model unchanged and places deviation from the linear reference in the cost function.Source: §1.1 · p. 2
1.2 Scope of this Work
- The paper extends Backman et al. (2012) to track multiple reference segments within the prediction horizon and implements the approach in a guidance system using the tractor’s TIM steering interface.It also includes custom path-segment handling.Source: §1.2 · p. 2
2. METHODS
The method extends cost-based NMPC path tracking to multiple reference segments within the prediction horizon. It combines a smoothed distance cost with runtime path handling and a real-time tractor implementation.
2.1 Nonlinear Model Predictive Control
- The controller uses a finite-horizon nonlinear optimization with updated feedback conditions, nonlinear dynamics, weighted reference error, and state and control bounds.Only part of the computed control sequence, typically its first input, is applied before the next cycle updates initial conditions, parameters, or constraints.Source: §2.1 · p. 2
2.2 System Model
- A continuous-time kinematic bicycle model adds steering-angle derivative as a state, while steering-angle derivative is the control input and speed is a manually controlled parameter.The speed is treated as constant during each controller iteration but may change between iterations; the kinematic model is chosen for real-time requirements and available CAN-bus parameters.Source: §2.2 · p. 2
2.3 Cost function
- Sigmoid-based activation combines multiple piecewise-linear reference segments in a distance cost, smoothing the state-dependent changes between references for Newton-type optimization.Segment selectors use progress and activation order to target one segment; the first and last selectors remain unclamped so a viable reference can persist beyond the path ends.Source: §2.3 · p. 3
- Finite sigmoid steepness can leave opposing segments with a zero-cost critical region; increasing k sharpens the approximation but moves that region closer to the path.For the example, k = 3 leaves critical zones sufficiently far from the relevant tracking region, whereas k = 20 can lead to diverging control behavior.Source: §2.3 · p. 3, Figure 2 · p. 3
- The objective adds steering-angle-derivative cost to trade actuator effort against smoother motion with fewer control actions or faster error correction.The steering-rate term multiplies its corresponding weight from W by the steering actuator’s rate of change.Source: §2.3 · p. 3
2.4 Runtime Parameter Updating
- The path handler initializes the planned-path subset and updates viable adjacent reference segments each cycle using vehicle position and orientation.It filters short segments, creates a feasible startup connection, and limits startup curvature; insufficient points are handled by extending the last segment, while turns over 60° temporarily replace subsequent segments with linear extensions.Source: §2.4 · p. 4
2.5 Real-time implementation
- The real-time implementation uses acados and HPIPM with a 100-step horizon at 100 ms, and controls tractor actuators through TIM over an ISO 11783 CAN bus.It predicts actuator-delay effects by propagating dynamics with prior commands and constant speed; prediction accuracy falls with speed variation or system delay, degrading tracking performance.Source: §2.5 · pp. 4–5
3. MATERIALS
The field-test tractor used a TIM steering interface and a dual-antenna GNSS setup for position and heading measurements.
3. MATERIALS
- The Fendt 314 Vario Gen 4 test tractor was equipped with a TIM interface capable of controlling its steering system.The steering angle came from TIM messages, while speed came via the standardized ISO 11783-7 (2022) message.Source: §3 · p. 5
- Two roof-mounted GNSS antennas, separated by 1.625 m, fed a receiver with RTK corrections from an on-site reference station.The main antenna position was estimated accurate to 2 cm, and the reported heading-angle standard deviation was 0.15°.Source: §3 · p. 5
- Steering angle was obtained from TIM messages, and speed from the standardized ISO 11783-7 (2022) message.Custom message-handling, TIM-client, and interprocess-communication routines were implemented but not further discussed.Source: §3 · p. 5
4. RESULTS
A grassland figure-eight trial evaluated tracking over varying curvature and measured NMPC solution times. The controller tracked the path with a 6.1 cm mean absolute cross-track error and converged within the anticipated control cycle.
4. RESULTS
- The controller was tested on a grassland figure-eight path with varying curvature and automatically generated start and end segments.The main accuracy metric was cross-track error; when it was undefined, the closest distance to path nodes was used instead.Source: §4 · p. 5, §4.1 · p. 5
4.1 Tracking Accuracy
- 6.1 cm mean absolute cross-track error; signed mean error was 0.2 cm, with an 8.4 cm standard deviation.Errors reached −26.5 cm and 32.9 cm; oscillations were partly attributed to linear segments and smoothing residuals, while temporary roll without IMU-based projection may prompt unnecessary corrections over bumps.Source: Figure 5 · p. 5, §4.1 · p. 5
4.2 NMPC Solution Statistics
- 3.45 ms average solve time (9.0 ms maximum); convergence held throughout, with a median of 2 and maximum of 5 SQP iterations.Execution remained below the anticipated 100 ms control cycle despite the 100-step prediction horizon.Source: Figure 6 · p. 6, §4.2 · p. 6
5. CONCLUSION
The work extends tractor NMPC path tracking to multiple linear reference segments and introduces a method for selecting segments from a predefined path.
5. CONCLUSION
- The controller incorporates multiple linear reference segments into the NMPC objective, enabling longer prediction horizons than previous implementations.A method for selecting the required segments from a predefined path was also introduced.Source: §5 · p. 6
- The controller was successfully tested in a field experiment using the tractor’s TIM steering interface.The evaluated setup did not match the accuracy or refinement of commercial integrated guidance systems.Source: §5 · p. 6
- The authors describe the real-time capability of the NMPC controller, despite its long horizon and extensive cost formulation, as promising.They suggest future application to more challenging systems may enable use cases not viable with current guidance systems.Source: §5 · p. 6