Before sensor fusion
Summary: modern navigation addresses the need for continuity between the robot, its local surroundings and a global mission.
Before modern localization stacks, agricultural automation mostly relied on repeated paths, human operators, mechanical guidance or highly structured environmental references.
These approaches are adequate for simple operations, but become fragile when a robot has to navigate an unknown environment, avoid dynamic obstacles and operate in regions where one or more sensors are unreliable. The central problem is pose continuity: the robot must know where it is, in which frame, and with what confidence.
Any human intervention degrades the potential productivity and economic gains of automation; therefore a more robust navigation algorithm is desirable.
Frames and alignment
Summary: an agricultural robot must coordinate several different reference frames from different sensors.
A coordinate frame is a way to describe position and orientation relative to a known point. A robot usually has a base_link frame attached to its chassis, an odom frame for short-term motion, a map_local frame relative to nearby features, for example the walls of the room the robot is in, and a global map_global_ENU frame tied to GNSS systems such as GPS.
The problem appears when the local map and global frame do not point in exactly the same direction or do not share the same origin. If this offset is ignored, the robot can appear to “jump” when GNSS information becomes available again, for example when exiting a building. This problem cannot be ignored by trusting only one sensor, because none of them alone are suited for fully autonomous use.
Comparison of sensors
Summary: GNSS is suited for global positioning, LiDAR and cameras for local positioning, and IMU and encoders for instantaneous movement.
The following table provides a short comparison of commonly used sensors.
| Source | Strengths | Limits | Best environments |
|---|
GNSS-RTK GPS and similar systems | Global reference frame shared across all GNSS users, up to centimeter-level accuracy when corrected with RTK, independent of local features, low cost and small size. | Physical occlusion such as canopy and buildings, multipath effects and loss of RTK signals. | Open fields, long trajectories, returning to a plot. |
LiDAR Light Detection and Ranging | High-density and accurate information about nearby objects, more robust to lighting conditions than RGB cameras. | Frame can drift if the environment is repetitive or featureless; still vulnerable to dust and rain; equipment cost. | Spaces with many obstacles, such as garages and greenhouses. |
IMU Inertial Measurement Unit | High update frequency, low cost and small size. | Drifts over time, sensitive to mounting and vibration. | Tracking short-term movement and orientation changes. |
Rotational encoders Wheel spin | High update frequency, low cost and small size, physically coupled to the measured surface. | Wheel slip; relies on accurate information about wheel or track dimensions. | Firm soil, low speed. |
| RGB/stereo/depth cameras | Versatile: they not only detect features, but can classify them. | Sensitive to environmental conditions such as lighting, weather and lens cleanliness. Requires calibrated vision models and enough computing resources. | Row tracking, crop detection, differentiating obstacles. |
While commonly used sensors cover the whole positioning scope from instantaneous transitions to global positioning, each scale level has disadvantages that are complemented by other sensors.
Sensor fusion
Summary: fusion turns imperfect measurements into a more stable estimate by weighing their credibility according to context.
The natural conclusion from the comparison is that no sensor is reliable everywhere, and higher precision can be achieved by fusing inputs from multiple sources. However, this fusion must take into account the reliability and precision of each sensor type, which vary when transitioning between environments.
The following are commonly used fusion methods. They are not mutually exclusive, and can be used together at different levels in the robot software. For example, IMU and wheel encoder inputs can be fused into a pose estimate with a Kalman filter, and a factor graph can then fuse that estimate with the history of GNSS measurements.
Filtering methods
Classic approach to sensor fusion. The most popular filters are Extended and Unscented Kalman filters (EKF/UKF), and particle filters. Generally speaking, these assume that the future position depends only on the present state, a hidden Markov process, and predict this future state using current observations and assumptions about each sensor error. This method's accuracy strongly depends on those assumptions, and may require fine-tuning.
Factor graphs
More modern approach. Instead of basing the prediction only on the previous state, factor graphs encode pose estimates as nodes connected to adjacent ones by sensor measurements, then solve for the set of estimates, in other words the robot path through time, that best matches the history of observations. Managing storage and pruning of past measurements then becomes an issue.
Neural networks
Commonly known as AI. Neural networks use a large series of simple operations to approximate an unknown function. While it is theoretically possible to automate the entire task using a single neural network, they require extensive training and their behavior in unseen situations is not well understood.
Theoretical workflow: from local to global
Summary: the transition builds a local estimate, waits for reliable GNSS, then progressively computes the link between local map and global coordinates.
The following is an example workflow of how a robot can transition from a GNSS-denied environment to an open one, and handle the new information received.
Define frames
Document base_link, odom, map_local, map_global_ENU, lidar, imu, camera and gnss. Axis conventions must be explicit.
Create local localization
Use LiDAR, IMU and odometry to build a map or localize the robot in the garage and GNSS-denied areas.
Detect reliable GNSS return
Monitor RTK fix/float, covariance, GDOP, satellite count and dual-antenna heading coherence.
Align the local map with ENU
Once a reliable GNSS signal is obtained, use global position and initial heading to compute the transform between map_local and map_global_ENU.
Insert an intermediate frame
A map_local_aligned frame limits pose jumps and supports a gradual shift between local and global sources.
Activate adaptive fusion
Increase GNSS weighting when it is stable, and return to LiDAR/IMU/vision when the signal degrades.