In the landscape of mobile robotics, localization is not merely an answer to the question ‘where am I?’; rather, it forms the backbone of an autonomous system’s operational awareness. Without a precise and continuous position estimation, many high-level functions – from path planning to dynamic obstacle avoidance – drastically lose their effectiveness. This ultimately compromises the entire mission.
The Limitations of GNSS Alone in Agriculture
For decades, GNSS (Global Navigation Satellite System) has been considered the most advanced tool for assisted and autonomous driving in agriculture. However, practical application in complex scenarios has highlighted insurmountable structural limitations.
In environments characterized by dense vegetation, metallic structures, or hilly terrains, satellite signals suffer critical distortions due to multipath effects or shadowing phenomena. These issues are not merely technical glitches; in fact, they translate into:
- Unplanned downtime: Interruption of critical crop protection (phytosanitary) treatments.
- Increased costs: The need for emergency manual labor and the potential risk of crop loss.
- Dynamic instability: Discontinuous position estimation can lead the control system to make abrupt and sudden trajectory corrections.
- Safety risks: Sudden trajectory corrections can damage the soil, the crops, or, even worse, endanger operators working alongside the robot.
LiDAR Odometry: The New Frontier of Navigation
A concrete solution to these critical challenges is represented by LiDAR Odometry (LO). This technology estimates the robot’s movement by analyzing the stream of data coming from the LiDAR sensor.
The integration of an Inertial Measurement Unit (IMU), which leads to so-called LiDAR Inertial Odometry (LIO) algorithms, is often crucial for eliminating motion distortion (deskewing) from the LiDAR point cloud and providing a solid inertial reference during rapid movements of the vehicle. This synergy ensures excellent support for position estimation even in geometry-poor environments. Furthermore, it guarantees the spatial consistency of the data where LiDAR alone would lose accuracy due to a lack of environmental features. From this point forward, for the sake of simplicity, we will refer to ‘LiDAR Odometry’ to generically indicate both LO and LIO algorithms.
The Open Field Paradox
Despite its effectiveness, LiDAR Odometry is not without its challenges. Because it inherently relies on the structure of the surrounding environment, LiDAR Odometry algorithms struggle in scenarios devoid of distinct geometric features. In an open, flat field, the scarcity of prominent landmarks prevents the sensor from accurately determining the movement made.
Performance analysis
LiDAR Odometry has garnered significant interest in recent years. Consequently, various solutions are currently available and ready for integration into mobile robots, each with its own distinct characteristics. At Aitronik, we evaluated four of the most prominent open-source LiDAR Odometry tools: KISS-ICP, LIO-SAM, GLIM, e RKO. To assess the robustness of these four algorithms, we captured point cloud data using a SICK Multiscan 165 LiDAR. The LiDAR was integrated onboard the ASNOT ground robot, which was deployed in autonomous navigation along a path encompassing both dirt roads and agricultural areas.
During its path, the robot encountered both areas with a scarcity of distinctive features (left image, where the lack of vegetation and objects within the LiDAR’s field of view prevents the acquisition of salient spatial landmarks) and zones much richer in features (right image). In the former case, the generated point cloud can be so devoid of information that, consequently, the output of the LO algorithms can become highly inaccurate.
Evaluation of the Four Tested Algorithms
This becomes particularly evident during robot rotations. In Figure 4, the angular deviations between the trajectories estimated by the algorithms and the ground truth are primarily due to robot rotations occurring in feature-poor areas.
From the analysis of Figure 4, it also emerges that nearly all standard odometry systems provide excellent estimations in feature-rich areas. However, they lose consistency as soon as the robot traverses zones with scarce geometric characteristics. This leads to a general degradation of the estimation, as highlighted in Figure 4. The incorrect orientation of certain trajectories relative to the reference path is clearly visible in the segment following the feature-poor area.
The RKO algorithm
During the tests conducted, the RKO algorithm proved to be the top-performing one. Figure 5 shows the deviation between the estimates provided by RKO and the path actually traveled (indicated by the dashed line in Figure 5).
Although the algorithm provides a good estimation of the path, at certain points (highlighted by the brighter colors in Figure 4, where the error is greater), the lack of environmental features detectable by the LiDAR leads to significant errors in the trajectory estimation. Since LiDAR Odometry is not assisted by anchors or LiDAR georeferencing, these errors cannot be recovered. Conversely, knowing the position of fixed landmarks (e.g., optical markers) within the environment enables the automatic correction of the errors introduced by LiDAR Odometry, leading to a substantial improvement in localization.
Relative Position Error (RPE) and Operational Environment
Our tests on the evaluated algorithms show a direct correlation between accuracy and context:
- Structured Environments: In areas rich in buildings or dense vegetation, the relative position error (RPE) approaches zero (represented in blue in Figure 5), thanks to the ease with which the LiDAR recognizes landmarks.
- Open Fields: In the zones highlighted in red (Figure 5) and corresponding to the error peaks in the graph over time (Figure 6), the algorithm struggles to maintain accuracy in the absence of solid reference points.
Conclusions: towards sensor fusion
LiDAR Odometry is an important tool, but to guarantee the robustness required for 24/7 operation under all conditions, it cannot act alone. Therefore, the winning strategy lies in sensor fusion, that is the integration of multiple devices (such as LiDAR, GNSS, and IMU) to compensate for their reciprocal limitations.
Riferimenti ai tool di LIDAR Odometry utilizzati
| Algorithm | Source code | Paper |
| KISS-ICP | KISS-ICP su github | KISS-ICP su Arxiv |
| LIO-SAM | LIO-SAM su github | LIO-SAM su Arxiv |
| GLIM | GLIM su github | GLIM su Arxiv |
| RKO | RKO su github | RKO su Arxiv |
