⚡ electrical engineering

Resilient Navigation for Autonomous Farm Robots by Leveraging Jerk-Augmented Models with IMU-Only Disturbance Rejection

This paper presents a robust navigation algorithm for autonomous farm robots that combines a jerk-augmented Extended Kalman Filter with a Multiple Tuning Factor adaptation method to dynamically reject IMU-only disturbances and sensor outages, significantly improving 3D position accuracy in off-road environments.

Original authors: Batu Candan, Mohammed Atallah, Simone Servadio, Saeed Arabi

Published 2026-06-10

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Batu Candan, Mohammed Atallah, Simone Servadio, Saeed Arabi

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 drive a remote-controlled car through a bumpy, muddy field. You have a GPS to tell you where you are, but the signal keeps getting blocked by trees, and the car is shaking so violently from the bumps that its internal "feeling" sensors (accelerometers) are getting confused. It's like trying to read a map while someone is shaking the table you're sitting on.

This paper presents a new way to help farm robots navigate these messy conditions without getting lost. Here is the simple breakdown of how they did it:

The Problem: The "Shaky Table" Effect

Standard navigation systems for robots usually assume that if a car is moving, it's moving smoothly. They expect the car to either keep going at the same speed or slow down/accelerate gently.

But in a farm field, the ground is full of ruts, rocks, and mud. When the robot hits a bump, it gets a sudden, violent jolt.

The Old Way: The robot's computer thinks, "Whoa, I just accelerated super fast! I must be moving incredibly quickly!" It panics and updates its position based on that false jolt. Then, when the jolt stops, the robot is convinced it's in the wrong place. This causes it to drift off course.
The Sensor Issue: The robot's GPS might also glitch or disappear behind a tree. If the robot is relying on shaky sensor data, it needs a way to know, "Hey, this data point is garbage, ignore it," without throwing away all the good data.

The Solution: Two New Tricks

The researchers built a smarter navigation system using two main ideas:

1. The "Jerk" Factor (Predicting the Bump)

In physics, "Jerk" isn't about being rude; it's the technical term for how quickly acceleration changes.

The Analogy: Imagine you are driving a car.
- Velocity is how fast you are going.
- Acceleration is how hard you are pressing the gas pedal.
- Jerk is how suddenly you slam your foot down on the gas or hit the brakes.
The Fix: The old robot models only looked at speed and acceleration. They were like a driver who expects the road to be smooth. The new model adds "Jerk" to its brain. It anticipates that the robot might get hit by a sudden bump. By expecting the "suddenness" of the shake, the robot doesn't get fooled by it. It knows, "I'm shaking, but I'm not actually speeding up that fast."

2. The "Smart Filter" (Ignoring Only the Bad Data)

The robot has sensors that measure movement in three directions: Up/Down, Left/Right, and Forward/Backward.

The Old Way: If the robot hit a deep rut and the "Up/Down" sensor went crazy, the old systems would say, "Everything is wrong!" and throw away all the sensor data (including the good Left/Right and Forward/Backward data). It's like throwing out a whole basket of apples because one is bruised.
The New Way (Multiple Tuning Factor): The new system is like a very careful inspector. It looks at the "Up/Down" sensor and sees, "That's a bruised apple." It ignores only the Up/Down data. But it keeps listening to the Left/Right and Forward/Backward sensors because they are still working perfectly. This allows the robot to stay on course even when one part of its body is getting hammered by the terrain.

The Results: A Smoother Ride

The team tested this on a real farm robot called the Salin247, which looks like a small, electric, four-wheeled tank designed to drive through crops without crushing them.

They drove the robot through a real field with uneven ground and compared the new system against the old standard system.

The Old System: The robot's path looked jagged and drifted away from where it actually was, especially when hitting bumps or turning.
The New System: The robot stayed glued to its intended path. Even when the ground was shaking violently, the robot knew where it was.

The paper claims that this new method reduced the robot's position errors by more than ten times compared to the old method. It essentially turned a robot that was stumbling and guessing into one that could walk confidently through a storm.

Summary

Think of this paper as teaching a robot how to drive on a rollercoaster. Instead of getting dizzy and falling off, the robot learned to:

Expect the sudden drops (by understanding "Jerk").
Ignore the specific part of the ride that is shaking while trusting the parts that are steady (by using the "Smart Filter").

This allows farm robots to work reliably in the messy, bumpy real world, rather than just in perfect, smooth test labs.

Technical Summary: Resilient Navigation for Autonomous Farm Robots

Problem Statement
Autonomous agricultural robots operate in complex, unstructured off-road environments where precise navigation is frequently compromised by two primary factors: sensor outages (GNSS, LiDAR, visual) and high-frequency mechanical vibrations. Standard navigation approaches, typically relying on 6-state or 9-state Extended Kalman Filters (EKF) with assumptions of constant velocity or acceleration, struggle in these conditions. These models suffer from lag and cumulative errors when robots encounter impulsive terrain disturbances or sudden torque changes because they lack the higher-order derivative (jerk) necessary to model the rate of change of acceleration. Furthermore, existing adaptive filtering methods often employ scalar tuning factors that scale the entire measurement covariance matrix uniformly. This "all-or-nothing" approach is sub-optimal for multi-axis systems; for instance, a vertical impact from a soil rut may corrupt vertical accelerometer data while lateral data remains valid. Uniformly scaling the covariance matrix forces the filter to discard useful information from unaffected axes, reducing overall precision.

Methodology
The paper proposes a resilient navigation framework integrating two core innovations: a Jerk-Augmented State Space and a Multiple Tuning Factor (MTF) adaptation method within an EKF architecture.

Jerk-Augmented EKF: The state vector is expanded to include 3D position, velocity, and jerk (the rate of change of acceleration). By modeling jerk as a random walk process, the filter can account for the dynamics of sudden terrain-induced shocks and uneven soil densities. The state transition matrix is derived from a third-order Taylor expansion, allowing the system to predict the rate of change of acceleration rather than assuming it remains constant.
Multiple Tuning Factor (MTF) Adaptation: Instead of using a scalar to scale the entire measurement noise covariance matrix ( $R$ ), the proposed method dynamically adjusts the covariance along each axis independently. The algorithm monitors the innovation sequence (residuals) at each time step. If the trace of the squared residual exceeds a theoretical threshold, indicating an external disturbance, the algorithm calculates a tuning matrix ( $S_t$ ). This matrix inflates the covariance specifically for the axes experiencing high residuals (e.g., the vertical axis during a bump) while leaving the covariance of unaffected axes (e.g., lateral) unchanged. This selective isolation prevents the filter from discarding valid sensory data from non-impacted axes.
Attitude Representation: To handle steep maneuvers on uneven terrain without singularities, the system utilizes a Direct Cosine Matrix (DCM) for attitude representation rather than Euler angles or quaternions, extracting roll and pitch for gravity compensation.

Key Contributions

State Augmentation: Introduction of jerk into the EKF state vector to model the rate of change of acceleration, addressing the lag inherent in constant-acceleration models during impulsive disturbances.
Axis-Specific Adaptation: Development of an MTF adaptation strategy that dynamically adjusts measurement noise covariance on a per-axis basis. This allows the system to isolate localized disturbances without degrading the estimation of unaffected axes.
Robust Framework: A unified navigation algorithm that combines jerk modeling with adaptive covariance tuning to maintain precision in environments with high-frequency vibrations and sensor outages.

Experimental Results
The algorithm was validated using real-world data from a Salin247 autonomous agricultural robot (a 4-wheel independent drive electric platform) operating in unstructured farm environments. The proposed Jerk-Augmented MTF-EKF was compared against a baseline 6-state (Position-Velocity) EKF.

Error Reduction: The results demonstrate that the proposed method significantly reduces the 3D position Root Mean Square Error (RMSE) compared to the baseline. Specifically, the paper notes a reduction of more than ten times in 3D position RMSE.
Disturbance Rejection: While the baseline EKF exhibited "spikes" in position error and trajectory drift when encountering surface irregularities (e.g., at $t=45s$ and $t=90s$ ), the jerk-augmented filter successfully absorbed impulsive forces. This resulted in a smoother trajectory that adhered closely to the RTK ground truth.
Attitude Stability: The system maintained stable tracking of roll and pitch angles (RMSE of 0.0051 rad and 0.0076 rad, respectively) even during high-vibration intervals, confirming the effectiveness of the gravity compensation mechanism.

Significance
The paper claims that accounting for the rate of change of acceleration (jerk) is critical for minimizing lag and drift during dynamic maneuvers in agricultural settings. By combining higher-order kinematic modeling with axis-specific noise adaptation, the proposed framework provides superior dead-reckoning capabilities when GNSS or other sensors are degraded or when the robot encounters unpredictable terrain. The authors conclude that this approach enables robust, centimeter-level positional accuracy, which is essential for precision farming operations such as weeding, spraying, and harvesting in unstructured environments. Future work is identified as integrating visual odometry cues into the MTF framework to further enhance resilience during extended sensor outages.