Technical Summary: A Predictive Control Strategy to Offset-Point Tracking for Agricultural Mobile Robots

Problem Statement

Agricultural robotics is increasingly vital for sustainable farming, yet existing path-following controllers predominantly focus on the robot's center of motion (rear axle). This approach neglects the spatial footprint and dynamics of attached implements (e.g., weeders, cultivators), which are often rigidly mounted and directly interact with crops and soil. In practice, ignoring the implement's position can degrade tracking performance and cause crop damage, particularly during curvature transitions.

The core challenge addressed is the control of an offset point (the implement-soil interaction point) on a rigidly linked agricultural vehicle. Unlike trailer systems with pivot joints that naturally reduce tracking errors during turns, rigidly mounted implements create a significant lever arm. This configuration introduces kinematic constraints where the reference trajectory for the implement may be physically infeasible (e.g., requiring curvature discontinuities the vehicle cannot execute). Furthermore, agricultural environments introduce lateral slip (sideslip) due to soil conditions, which traditional models often ignore. The goal is to design a control strategy that minimizes unavoidable tracking errors while ensuring operational safety and precision.

Methodology

The authors propose a closed-form predictive control strategy that extends a previous model [1] to account for lateral sliding and rigid linkages. The methodology is structured as follows:

1. Kinematic Modeling with Sideslip

The system is modeled using an Ackermann-type bicycle model extended to include sideslip angles ($\beta_R$ and $\beta_F$) for rear and front wheels.

• Offset Point Definition: The implement point $I$ is defined by local coordinates $(I_s, I_y)$ relative to the robot frame.
• Tracking Error: The implement tracking error $e_I$ is formulated as a function of the robot's lateral error $y$, angular deviation $\tilde{\psi}$, and the geometric offset.
• Sliding Observer: Since sideslip angles are not directly measurable, an observer (based on [37]) estimates $\beta_R$ and $\beta_F$ using GNSS-RTK and IMU data to compute lateral velocity disturbances.

2. Predictive Control Framework

The control strategy employs a two-stage backstepping approach enhanced with a predictive horizon:

• Stage 1 (Desired Orientation): Instead of a purely reactive correction, the method predicts the tracking error evolution over a horizon $s_h$. It defines an optimization criterion $J$ that minimizes the cumulative squared error between the predicted implement trajectory and a desired convergence profile $h(s)$. This profile accounts for the specific kinematic behavior where rear-mounted implements must initially diverge from the path before converging (lever-arm effect).
• Closed-Form Solution: By approximating the error dynamics and neglecting higher-order curvature terms (valid for typical agricultural paths), the authors derive a closed-form solution for the optimal desired angular deviation $\tilde{\psi}_d$. This avoids the computational burden and reliability issues of numerical solvers used in standard Model Predictive Control (MPC).
• Stage 2 (Steering Control): A secondary backstepping loop adjusts the steering angle $\delta$ to drive the robot's actual orientation toward the desired $\tilde{\psi}_d$, ensuring exponential convergence.

Key Contributions

1. Extended Model with Lateral Slip: The paper extends the offset-point control model from [1] to explicitly incorporate lateral sliding effects, which are critical in agricultural environments.
2. Novel Predictive Control Strategy: A new closed-form predictive controller is introduced for rigidly linked implements. Unlike standard MPC, it does not require numerical optimization, ensuring real-time feasibility and guaranteed convergence.
3. General Applicability: The method is designed to handle both front-mounted and rear-mounted implements, addressing the specific kinematic challenges (such as the "divergence before convergence" behavior) of rear-mounted configurations which existing predictive methods (e.g., [36]) fail to handle.

Experimental Results

The approach was validated through simulations and real-world field experiments on a four-wheel-drive Ackermann robot equipped with two implements: a mechanical weeder and a spring-tine cultivator.

Simulation Results

• Velocity and Offset: Simulations showed that tracking errors remain within acceptable limits (<15 cm median) across various speeds and offset positions. Rear-mounted implements exhibited slightly higher errors due to the lever-arm effect, necessitating longer prediction horizons at higher speeds.
• Parameter Sensitivity: The optimal prediction horizon was found to be robust, with performance remaining near-optimal over a range of values, reducing the need for extensive tuning.

Real-World Experiments

Experiments were conducted on grass and cultivated soil, comparing the proposed method against:

1. Lateral Servoing (reactive, geometric) [1].
2. Backstepping Control (reactive) [1].
3. MPC for Front-Mounted Implements [36].

Key Findings:

• Error Reduction: The proposed method reduced the median tracking error by 24% to 56% compared to state-of-the-art baselines.
• Peak Error Mitigation: During curvature transitions and obstacle avoidance maneuvers, peak errors were reduced by up to 70%.
• Rear-Mounted Performance: While the MPC method from [36] failed to converge for rear-mounted implements, the proposed method successfully tracked the path.
• Impact of Sliding Observer: Ablation studies confirmed that the sliding observer is critical, reducing median tracking errors by 50% to 64% in curved sections where soil slip is significant.
• Obstacle Avoidance: In a scenario involving obstacle avoidance with sharp deviations, the proposed method maintained tracking errors below 0.5 m, whereas reactive methods resulted in errors exceeding 1.0 m, posing a risk of crop damage or collision.

Significance and Claims

The paper claims that the proposed strategy offers a significant advancement in agricultural robotics by shifting the control focus from the vehicle center to the implement itself. The significance lies in:

• Safety and Precision: By minimizing tracking errors during curvature transitions, the method reduces the risk of damaging crops, which is critical for delicate operations like inter-row weeding.
• Robustness: The explicit modeling of lateral slip and the use of a predictive horizon allow the system to anticipate and compensate for disturbances that reactive controllers cannot handle.
• Practicality: The closed-form nature of the controller ensures low computational cost and high reliability, making it suitable for deployment on embedded agricultural systems without the risk of solver failures common in numerical MPC.

The authors conclude that while perfect tracking is often physically impossible for rigid offset points due to kinematic constraints, their method effectively minimizes these unavoidable deviations, thereby enhancing the operational safety and efficiency of autonomous agricultural systems. Future work is directed toward extending the framework to skid-steered platforms and active implements with additional degrees of freedom.