TL;DR
This paper introduces LOCUS 2.0, a lidar odometry system designed for resource-constrained robots operating in challenging, GPS-denied environments like underground tunnels. It achieves significant improvements in computational efficiency and memory usage through novel GICP formulation, adaptive filtering, and a sliding-window map, making real-time 3D mapping feasible for heterogeneous robotic teams.
What Problem Does This Paper Address?
Autonomous robots often need to navigate and map complex environments where GPS signals are unavailable, such as underground mines, caves, or disaster sites. Lidar odometry is a promising solution for this, offering robust 3D sensing. However, existing lidar odometry systems can be computationally intensive and memory-hungry, making them unsuitable for deployment on smaller, less powerful robots or for large-scale mapping tasks where memory quickly becomes a bottleneck. This paper tackles the challenge of creating a lidar odometry system that is both accurate and efficient enough to run in real-time on heterogeneous robotic platforms with limited onboard computation and memory.
Methods
LOCUS 2.0 builds upon its predecessor, LOCUS 1.0, with several key innovations:
- Normals-based Generalized Iterative Closest Point (GICP): Instead of recalculating point covariances (a computationally expensive step) for GICP, LOCUS 2.0 leverages pre-computed point normals. This novel formulation significantly reduces the computation time for point cloud alignment, especially for scan-to-submap matching, by allowing covariances to be computed once and reused.
- Adaptive Voxel Grid Filter: To ensure a consistent computational load regardless of the environment's complexity or the number of lidar sensors, an adaptive voxel grid filter is employed. This filter dynamically adjusts its voxel size to maintain a fixed number of points, thereby guaranteeing a predictable processing time per scan.
- Sliding-Window Map: To address memory limitations in large-scale explorations, LOCUS 2.0 replaces the global map storage of its predecessor with a sliding-window approach. This method keeps only a local, robot-centered portion of the map in memory, dynamically updating it as the robot moves. The paper evaluates two implementations: a multi-threaded octree and an incremental k-d tree (ikd-tree).
Datasets
The paper introduces a new, large-scale lidar odometry dataset collected in challenging underground environments. This dataset features:
- Environments: Limestone mines, urban subway tunnels, and lava tubes.
- Conditions: Fog, dust, darkness, and geometrically degenerate areas.
- Platforms: Heterogeneous robotic platforms, including the legged Spot robot and the wheeled Husky robot.
- Duration: A total of 11 hours of operation.
- Distance: 16 km of traveled distance.
This dataset is made publicly available to foster further research.
Key Insights and Ideas
- Normals are a powerful shortcut: Using point normals to approximate covariance in GICP is a clever way to drastically reduce computation without a significant loss in accuracy. This is particularly impactful for scan-to-submap registration.
- Predictable performance is crucial: The adaptive voxel grid filter is key to ensuring that the system's performance (CPU load, latency) remains consistent, even when the environment or sensor setup changes. This is vital for real-time control and planning.
- Memory is the enemy of scale: The sliding-window map approach is a pragmatic solution for large-scale mapping, demonstrating that global maps are often infeasible for resource-constrained robots.
- Real-world challenges are diverse: The newly released dataset highlights the variety of difficult conditions (e.g., feature-poor corridors, poor lighting, dynamic terrain) that lidar odometry systems must contend with in real-world underground exploration.
Key Results
- Computational Efficiency: The normals-based GICP formulation reduces computational metrics (CPU usage, odometry delay, scan-to-scan/submap times) by an average of 18.57% compared to standard GICP.
- Accuracy: While there's a slight increase in mean and max Absolute Pose Error (APE) (around 10.82%), this is considered a reasonable trade-off for the significant computational gains. Excluding one challenging tunnel dataset, the APE increase drops to 5.23%.
- Memory Reduction: The sliding-window map approaches (ikd-tree and multi-threaded octree) achieve substantial memory savings, with ikd-tree reducing memory usage by up to 68.09% compared to the previous global octree approach.
- Robustness: LOCUS 2.0 demonstrated robust performance across various challenging underground environments, including those where other state-of-the-art methods failed (e.g., lidar slip in tunnels).
- State-of-the-Art Comparison: LOCUS 2.0 achieves competitive or superior performance compared to FAST-LIO and LINS in terms of APE error across multiple datasets and environments.
Limitations and Future Work
- APE Increase: The paper acknowledges a slight increase in APE error with the normals-based GICP. Future work could explore methods to mitigate this while retaining computational efficiency.
- Ikd-tree Performance: The ikd-tree implementation, while effective for memory management, shows higher computation times for insertion, deletion, and search compared to the octree. Further optimization of the ikd-tree or exploration of alternative data structures could be beneficial.
- Sensor Fusion: While LOCUS 2.0 is designed for heterogeneous platforms, the paper primarily focuses on lidar odometry. Integrating other sensors more deeply (e.g., visual odometry, inertial measurement units) could further enhance robustness and accuracy, especially in extremely feature-poor environments.
- Real-time Performance on Diverse Hardware: While the system is designed for resource-constrained platforms, further extensive testing and optimization across a wider range of hardware specifications would be valuable.
Summary updated