This paper, recently published in IEEE Robotics and Automation Letters, introduces GRIL-Calib, a targetless extrinsic calibration method specifically designed for ground robots. It tackles the "planar motion" problem that often causes other calibration tools to fail.
TL;DR
GRIL-Calib enables precise 6-DoF IMU-LiDAR calibration for ground robots by using the ground plane itself as a geometric constraint. By switching from sequential to simultaneous optimization, it solves for "unobservable" parameters (like vertical offset) that typically plague robots restricted to 2D planar motion.
What Problem Does This Paper Address?
For an IMU and LiDAR to work together (sensor fusion), the system must know their exact relative position and orientation—this is extrinsic calibration.
Most "targetless" methods (which don't require a special checkerboard or room) rely on "full excitation." This means you have to move the sensors in all six degrees of freedom (tilting, rotating, and moving in every direction) to make the calibration parameters mathematically "observable."
The catch: Ground robots are stuck on the floor. Their motion is mostly planar. Because they don't tilt or jump, certain calibration parameters—specifically the vertical translation ($z$-axis) and some rotation angles—become "degenerate" or unobservable. Existing methods often exhibit significant drift or fail entirely in these scenarios.
Methods
The authors propose three main innovations to overcome planar degeneracy:
- Ground-Aware LiDAR Odometry (LO): They enhance standard LiDAR odometry by incorporating ground plane residuals into an Iterative Error State Kalman Filter (IESKF). This forces the odometry to respect the fact that the robot is on a 2D plane, significantly reducing $z$-axis drift.
- Ground Plane Motion (GPM) Constraint: This is the core novelty. By identifying the ground plane in the LiDAR data (using the Patchwork++ algorithm) and knowing the IMU's height from the ground as a prior, they create a geometric constraint. This allows the system to solve for the extrinsic parameters even in directions where the robot isn't moving.
- Single Optimization (SO) Process: Standard methods usually solve for rotation first, then translation. However, translation estimates are highly dependent on rotation. GRIL-Calib formulates the problem as a single, joint optimization, solving all 6-DoF parameters simultaneously to capture their correlations and improve accuracy.
Datasets
The researchers validated their method on three diverse, real-world datasets:
- M2DGR: A large-scale dataset for ground robots featuring various indoor and outdoor environments (gates, halls, elevators, and streets).
- S3E: A collaborative SLAM dataset used to test the method's consistency across different hardware setups.
- HILTI: A challenging benchmark dataset often used in professional SLAM competitions.
Key Insights and Ideas
- The Floor is a Feature: Instead of viewing planar motion as a limitation, GRIL-Calib treats the ground as a permanent, reliable geometric reference.
- Observing the Unobservable: Mathematically, you can't "see" a vertical offset if you only move horizontally. However, by linking the IMU's known height to the LiDAR's detected ground plane, that offset becomes solvable.
- Coupling Matters: The "Aha!" moment in their optimization strategy is that treating rotation and translation as independent steps loses information. Solving them together (the SO approach) led to a 51% improvement in rotation accuracy and a 23% improvement in translation accuracy compared to sequential methods.
Key Results
GRIL-Calib consistently outperformed state-of-the-art baselines like LI-Init and FAST-LIO2:
- Accuracy: On the M2DGR dataset, it achieved rotation errors as low as 0.501° and translation errors as low as 0.034m (3.4 cm).
- Robustness: The method proved far more resilient to poor "initial guesses." Even when the starting parameters were significantly off, the ground plane constraints "pulled" the optimization toward the correct solution.
- Ablation Success: Adding the GPM constraint alone improved translation accuracy by 70% compared to the baseline.
Limitations and Future Work
The primary limitation acknowledged is the flat ground assumption. The current GPM constraint relies on the ground being a relatively smooth plane. If a robot is operating on highly irregular, rocky, or non-planar terrain, the geometric constraints might introduce errors.
Future Work: The authors suggest extending the algorithm to handle non-flat terrains and potentially incorporating more complex environmental features to further relax motion requirements.
Code Availability: The authors have open-sourced their implementation at https://github.com/Taeyoung96/GRIL-Calib.
Summary updated