TL;DR
This paper introduces LAMP, a robust multi-robot SLAM system designed for challenging subterranean environments. It tackles issues like poor sensor conditions and repetitive features by employing an accurate lidar front-end and a flexible back-end that effectively rejects erroneous loop closures, enabling accurate mapping for exploration and disaster response.
What Problem Does This Paper Address?
Autonomous exploration of subterranean environments, such as tunnels, mines, and lava tubes, presents significant challenges for Simultaneous Localization and Mapping (SLAM) systems. These environments are often characterized by:
- Perceptually-Degraded Conditions: Poor lighting, dust, and water can render vision-based SLAM unreliable.
- Challenging Terrain: Uneven and slippery surfaces make wheel odometry inaccurate.
- Featureless Environments: Long, monotonous corridors lack distinct landmarks, leading to sensor drift and ambiguity.
- Perceptual Aliasing: Repetitive structures (e.g., similar-looking tunnels) can cause spurious loop closures, severely distorting the map.
These issues hinder the creation of accurate 3D maps, which are crucial for applications like disaster response, infrastructure monitoring, and the exploration of extraterrestrial worlds.
Methods
The LAMP system is a centralized multi-robot SLAM architecture with a focus on lidar-based mapping and robust loop closure handling. Key components include:
- Lidar Front-end:
- Lidar Odometry: Employs scan-to-scan matching using the Generalized Iterative Closest Point (GICP) algorithm for initial pose estimation. Crucially, it also incorporates scan-to-submap matching, which significantly improves accuracy by matching the current scan against a larger local map.
- Loop Closure Detection: Compares current scans with historical scans within a radius. GICP is used to find potential loop closures, with a fitness score threshold to filter out obviously incorrect matches.
- Vision Front-end: Utilizes YOLO for detecting "artifacts" (objects of interest) and the AprilTags library for fiducial markers, estimating their 3D positions using bearing and depth information.
- Back-end (Pose Graph Optimization - PGO):
- Local (Single-Robot) PGO: Each robot maintains its own pose graph, incorporating odometry and validated loop closures.
- Base Station (Multi-Robot) PGO: A central base station fuses pose graphs from all robots. It performs inter-robot loop closure detection and utilizes an Incremental Consistent Measurement Set Maximization (ICM) module to robustly reject outlier loop closures before optimizing the global pose graph.
- Human-in-the-loop Pose Graph Editing: Allows operators to manually identify and add loop closures, enhancing system flexibility.
Datasets
The paper evaluates the LAMP system in several real-world underground environments:
- Six underground tunnels and mines across the United States (listed in Table I).
- Specific mention of experiments conducted in the Bruceton Research mine and the Tunnel Circuit of the DARPA Subterranean Challenge.
- Ground truth data was provided by DARPA for artifact locations in the Bruceton mines. Robot trajectories were approximated by enforcing these ground truth artifact locations within the pose graph optimization.
Key Insights and Ideas
- Scan-to-Submap is Crucial for Odometry: The paper strongly demonstrates that scan-to-submap matching significantly outperforms simple scan-to-scan matching in terms of odometry accuracy, especially in environments with drift.
- Robust Loop Closure Rejection is Paramount: Perceptual aliasing in subterranean environments makes spurious loop closures a major problem. The ICM module is shown to be highly effective in preventing these false positives from distorting the overall map.
- Subterranean Environments are a Unique SLAM Challenge: The paper highlights how off-nominal sensor conditions, lack of features, and repetitive structures create a distinct set of problems not typically encountered in outdoor or well-lit indoor SLAM.
- Multi-robot SLAM Benefits from Centralized Robustness: While distributed SLAM is an area of research, this paper shows the effectiveness of a centralized approach when combined with robust outlier rejection for multi-robot coordination in complex environments.
Key Results
- Odometry Drift: Scan-to-submap odometry achieved a drift of approximately 1% of the distance traveled in nominal conditions in the Safety Research mine.
- Loop Closure Impact: Maps generated with loop closure detection showed significantly fewer spurious corridors and distortions compared to those without it.
- ICM Effectiveness: The ICM module dramatically reduced end-to-end drift. For instance, in one test, the drift was reduced from 2.8m to 0.2m by using ICM.
- Artifact Localization: The paper presents boxplots showing artifact localization accuracy, demonstrating improvements when loop closures and ICM are enabled.
- Multi-robot Mapping: Experiments with two robots showed that without loop closures, trajectories diverged significantly. With loop closures and ICM, the fused map was substantially improved, though a minor error was still observed in one instance.
Limitations and Future Work
- Lidar Odometry Limitations: In long, symmetric corridors without distinct features, lidar odometry can still underestimate robot motion, leading to drift.
- ICM Parameter Tuning: The ICM module requires extensive parameter tuning and can struggle to balance outlier rejection with inlier preservation due to approximate consistency checks.
- Scalability of Centralized SLAM: For more than four robots, communication bottlenecks can become an issue with the current centralized architecture.
- Future Work:
- Feature-based approaches for more robust global scan matching.
- Fusing information from other sensors (vision, IMU, wheel odometry) to improve odometry.
- Developing more robust global solvers for PGO that scale better and require less tuning.
- Exploring fully distributed SLAM solutions for better scalability.
- Investigating map compression techniques to reduce communication bandwidth in multi-robot systems.
- Developing automatic methods to adjust exchanged data based on available bandwidth.
- Researching compact and uncertainty-aware map representations for better risk prediction.
Summary updated