Rather than depending on a single positioning source, the researchers combined aerial and ground-based mapping.
High-resolution images of an apple orchard were first collected by drone and processed into a geographically referenced map. A ground robot then traveled through the orchard using LiDAR and inertial measurement sensors to record trees, terrain and other structural features from below the canopy.
Each sensing method fills a gap left by the other. Drone imagery can provide an accurate overhead reference when supported by global navigation satellite systems, but it reveals less detail beneath dense foliage. LiDAR gives a robot detailed three-dimensional information at ground level, yet its estimated position can gradually drift when satellite signals or other external references are limited.
To connect the two datasets, the researchers converted LiDAR measurements into bird’s-eye-view representations containing information on canopy height, structural density and surface reflectance. An AI model then compared those maps with the aerial imagery to identify corresponding orchard features, including tree rows, canopy formations and open ground.
The system uses separate feature-extraction networks for the aerial and LiDAR inputs, followed by transformer-based cross-attention and multi-scale processing. The goal is to determine where the robot’s ground-level observations fit within the geographically referenced drone map.
That match gives the navigation system an external reference point it can use to correct accumulated positioning errors.
During testing over approximately 1.3 kilometers of orchard travel, the researchers reported localization errors measured in centimeters and less long-distance drift than with conventional matching methods.
The approach was also tested under seasonal changes, which can alter foliage, shadows and overall canopy appearance. That is a relevant challenge for commercial orchards, where the same route can look substantially different throughout the growing cycle.
Researchers also designed the system with embedded computing in mind, potentially allowing mapping and localization functions to run onboard agricultural equipment rather than relying entirely on remote processing infrastructure.
Once the aerial and ground datasets are aligned, the information is incorporated into a pose-graph estimation framework to generate a geographically referenced, multi-layer model of the orchard.
For equipment manufacturers and agricultural technology developers, that type of digital map could have applications beyond simply keeping a robot on course.
The combined dataset can capture structural characteristics such as tree height and canopy information, creating a potential foundation for future crop or tree-health analysis. More reliable geographic positioning could also make it easier for autonomous equipment to return to specific trees, rows or treatment zones across repeated field operations.
Possible use cases include precision spraying, mowing, orchard inspection, crop transportation and robotic harvesting. In larger orchards, improved localization could also reduce the amount of manual route mapping or operator supervision required for autonomous systems.
Commercial readiness, however, has not yet been demonstrated across a broad range of farms. Further testing would be needed to assess how the technique performs with different crops, planting layouts, terrain, weather conditions and sensor configurations.
The research reflects a wider direction in agricultural robotics: combining multiple sources of field data rather than asking one sensor to handle every mapping and localization task.
For developers of autonomous farm equipment, the practical value may lie in creating a more consistent digital reference layer. By tying detailed ground-level sensing to geographically accurate aerial maps, robots could gain a clearer picture of both where they are and what is around them — two requirements that become increasingly important as farm automation moves toward more complex, repeatable field operations.
