TL;DR
This paper introduces LOCUS 2.0, a lidar odometry system designed for robust and efficient real-time 3D mapping in challenging, GNSS-denied environments. It tackles the critical issue of limited onboard computation and memory on robotic platforms by introducing novel GICP formulation, an adaptive voxel grid filter, and a sliding-window map approach.
What Problem Does This Paper Address?
Autonomous robots operating in environments like underground mines, caves, or disaster sites often lack reliable GPS signals, making localization a significant challenge. Lidar sensors are a promising solution, offering accurate 3D perception. However, existing lidar odometry systems can be computationally intensive and memory-hungry, hindering their deployment on resource-constrained robotic platforms, especially for large-scale mapping tasks. This paper aims to bridge this gap by developing a lidar odometry system that is both accurate and efficient enough to run in real-time on such platforms.
Methods
LOCUS 2.0 builds upon its predecessor, LOCUS 1.0, with several key improvements:
- GICP from Normals: A novel formulation of Generalized Iterative Closest Point (GICP) that leverages pre-computed point cloud normals. Instead of recalculating point covariances (a computationally expensive step), it uses normals to approximate them. This significantly speeds up the scan-to-scan and scan-to-submap registration process.
- Adaptive Voxel Grid Filter: This filter dynamically adjusts the voxel leaf size to maintain a fixed number of points for processing, regardless of the environment's geometry or the number of lidar sensors. This ensures a consistent and predictable computation load, crucial for real-time performance.
- Sliding-Window Map: To address memory limitations in large-scale environments, LOCUS 2.0 employs a sliding-window approach for map management. This keeps only a local, robot-centered portion of the map in memory, preventing unbounded memory growth. Two implementations are explored: a multi-threaded octree and an incremental k-d tree (ikd-tree).
- Health-Aware Sensor Integration: The system includes a health monitor that dynamically manages lidar inputs, ignoring delayed or failed sensors to maintain robustness.
Datasets
The authors introduce a new, challenging lidar odometry dataset collected from real-world underground environments (limestone mines, lava tubes, urban subway stations) during the DARPA Subterranean Challenge. This dataset features:
- Heterogeneous Platforms: Data collected from both legged (Spot) and wheeled (Husky) robots.
- Challenging Conditions: Includes environments with fog, dust, darkness, and geometrically degenerate features.
- Scale: A total of 11 hours of operation and 16 km of distance traveled.
- Ground Truth: Trajectories are derived from LOCUS 1.0 operating in a high-accuracy, scan-to-map mode against survey-grade maps.
This dataset is released as open-source to promote further research.
Key Insights and Ideas
- Normals as a Shortcut: The insight that point cloud normals can effectively approximate covariances for GICP is a significant computational win, especially for scan-to-submap matching where recalculating covariances for a large map would be prohibitive.
- Predictable Performance: The adaptive voxel grid filter is key to achieving consistent real-time performance. By decoupling computation from environmental complexity, it makes the system more reliable on resource-constrained robots.
- Memory Management is Crucial: The sliding-window map approach is a practical necessity for large-scale exploration. The paper demonstrates that even with increased computational cost for map updates, the memory savings are substantial and essential.
- Real-World Challenges: The new dataset highlights the importance of testing odometry systems in diverse and degraded conditions, which are common in real-world robotic applications.
Key Results
- Computational Efficiency: GICP from normals reduces computational metrics (CPU usage, odometry delay, scan-to-scan/submap times) by an average of 18.57% compared to standard GICP.
- Accuracy Trade-off: The GICP from normals approach shows a slight increase in mean and max Absolute Pose Error (APE) (around 10.82%), but this is considered a reasonable trade-off for the significant computational gains.
- Memory Reduction: The sliding-window map approaches (ikd-tree and multi-threaded octree) significantly reduce memory usage compared to a full map storage. For example, ikd-tree shows a -68.09% memory change relative to the baseline LOCUS 1.0.
- Robustness: LOCUS 2.0 demonstrates state-of-the-art performance across various environments, including challenging tunnel scenarios where other methods failed. It achieves the smallest APE errors in 5 out of 6 tested datasets when compared to FAST-LIO and LINS.
Limitations and Future Work
- APE Increase: While computational gains are substantial, there's a slight increase in APE. Future work could explore methods to mitigate this while retaining efficiency.
- Ikd-tree Performance: The ikd-tree data structure, while effective for memory management, incurs higher computation costs for search, deletion, and insertion compared to octree-based structures. Further optimization of ikd-tree operations could be beneficial.
- Sensor Fusion: While LOCUS 2.0 is designed to be a lidar-centric system, its architecture allows for the integration of other sensors. Future work could explore more tightly-coupled fusion strategies for enhanced robustness and accuracy.
- Dataset Expansion: The paper releases a valuable dataset, but continued collection and expansion of such challenging datasets will be crucial for advancing the field.
Summary updated