Abstract
Simultaneous Localization and Mapping (SLAM) is often considered a solved problem, yet most LiDAR-based algorithms remain tethered to specific hardware payloads. This lack of generalization complicates the deployment and maintenance of robot fleets. This paper introduces a unified 3D mapping framework that extends the Cartographer SLAM library to support diverse sensors, including fixed or rotating 2D and 3D LiDARs. By decoupling localization from sensor-specific data through a "pseudo-pose" interface, the system achieves high-quality real-time mapping across varied platforms with minimal parameter tuning.
The Challenge of Hardware Fragmentation
In practical robotics, SLAM is rarely "plug-and-play." Algorithms are typically optimized for a specific sensor's field of view (FOV), resolution, and data structure. For instance, transitioning from a drone equipped with a fixed 3D LiDAR to one using a rotating 2D scanner often necessitates a complete software overhaul. This fragmentation is particularly problematic in specialized environments like nuclear tunnels, where different scales of operation require different sensing payloads.
A Unified Localization Pipeline
To overcome these limitations, the authors replaced Cartographer’s standard localization module with a hybrid filtering approach. The core of this system is the interaction between an Error State Kalman Filter (ESKF) and a Gaussian Particle Filter (GPF).
The process works through a "Pseudo Pose" mechanism:
- Motion Modeling: The ESKF uses IMU data to provide a high-frequency pose prior.
- Particle Reweighting: The GPF takes this prior and samples particles, which are then reweighted based on how well the current laser scan matches the existing distance map.
- Pose Recovery: The GPF recovers a "pseudo pose" from the weighted particles.
- Update: This pseudo pose is fed back into the ESKF as a measurement.
Because the interface between the GPF and ESKF is a generic pose (position and orientation), the ESKF remains agnostic to whether the original data was 2D or 3D.
Hardware Agnosticism in Practice
The framework was validated on several distinct hardware configurations, demonstrating its versatility across different drone platforms and sensor types.
The researchers utilized a customized DJI M100 with a rotating Hokuyo 2D scanner and a DJI M600 with a rotating Velodyne VLP-16. They also tested fixed-mount configurations. A key technical hurdle for rotating sensors—the accumulation of 2D scans into 3D "frames"—was solved by grouping $N$ scans to form a point cloud that is then matched against local submaps.
Key Insights: FOV vs. Resolution
A significant finding of the study is that Field of View (FOV) often matters more than raw resolution for stability. In narrow indoor environments, fixed 3D LiDARs often suffer from limited vertical FOV, leading to "blind spots" at the floor or ceiling that cause vertical drift. By rotating a 2D or 3D sensor, the system achieves a much wider FOV, which significantly improves localization robustness in under-constrained spaces like long hallways.
Results and Performance
The system demonstrated impressive precision in both simulation and real-world tests:
- Simulation: In a digital twin of the DOE-PUREX nuclear tunnel, the framework achieved a translational drift rate of 1.34% over a 165m flight.
- Real-Time Execution: By leveraging multi-threading and efficient Octomap-based distance map updates, the system maintains real-time performance on standard ARM-based drone processors.
- Minimal Tuning: Switching between sensors primarily requires adjusting IMU noise parameters and map resolution, rather than rewriting core logic.
Limitations and Future Work
Despite its success, the framework faces challenges related to mechanical complexity and motion. Rotating sensors require high-precision synchronization and slip-rings to prevent wire tangling. Furthermore, high-speed flight can result in sparse scans that degrade localization stability. Future research will focus on improving motion compensation and adaptive sampling to support higher velocities.
Reference: Zhen, W., & Scherer, S. (2018). A Unified 3D Mapping Framework using a 3D or 2D LiDAR. arXiv:1810.12515v1 [cs.RO].
Summary updated