Introduction
Precise extrinsic calibration between an Inertial Measurement Unit (IMU) and a LiDAR is fundamental for robust sensor fusion in autonomous systems. While targetless calibration methods have gained popularity by removing the need for specialized calibration environments, they often struggle with ground robots. These robots typically operate in near-planar motion, leading to "degeneracy" where certain parameters—such as vertical translation and specific rotation angles—become mathematically unobservable.
The Planar Degeneracy Problem
Standard targetless methods require "full excitation," meaning the sensor suite must be moved through all six degrees of freedom (6-DoF). Ground robots, constrained to flat surfaces, cannot easily perform the tilting or vertical movements necessary to resolve all extrinsic parameters. This lack of motion leads to significant drift or failure in existing calibration tools.
Methodology: GRIL-Calib
To address these challenges, the authors of "GRIL-Calib" propose a novel approach that leverages the ground plane itself as a geometric constraint rather than a limitation.
Ground Plane Motion (GPM) Constraint
The core innovation is the Ground Plane Motion (GPM) constraint. By identifying the ground plane in LiDAR point clouds (using algorithms like Patchwork++) and utilizing the known height of the IMU from the ground as a prior, the system can resolve previously unobservable parameters like vertical offset.
Single Optimization (SO) Process
Unlike traditional sequential methods that solve for rotation and then translation, GRIL-Calib employs a Single Optimization (SO) process. This joint optimization solves all 6-DoF parameters simultaneously, capturing the intricate correlations between rotation and translation. This shift resulted in a 51% improvement in rotation accuracy and a 23% improvement in translation accuracy during testing.
Experimental Results
The method was validated across several high-profile datasets, including M2DGR, S3E, and HILTI. GRIL-Calib consistently outperformed state-of-the-art baselines like LI-Init and FAST-LIO2 in both accuracy and robustness.
On the M2DGR dataset, the system achieved rotation errors as low as 0.501° and translation errors within 3.4 cm. Furthermore, the GPM constraint proved vital for robustness, successfully "pulling" the optimization toward the correct solution even when provided with poor initial parameter guesses.
Conclusion and Future Work
GRIL-Calib demonstrates that the inherent constraints of ground robot motion can be turned into an advantage for calibration. By treating the floor as a permanent geometric feature, it achieves high-precision 6-DoF calibration without requiring complex manual excitation. While currently optimized for flat environments, future iterations aim to extend these constraints to irregular terrains, further broadening the applicability of targetless calibration for field robotics.
Code Availability: The implementation is open-sourced at https://github.com/Taeyoung96/GRIL-Calib.
Summary updated