Source-linked AI summary
LAMP: Large-Scale Autonomous Mapping and Positioning for Exploration of Perceptually-Degraded Subterranean Environments
Kamak Ebadi, Yun Chang, Matteo Palieri, Alex Stephens, Alex Hatteland, Eric Heiden, Abhishek Thakur, Nobuhiro Funabiki, Benjamin Morrell, Sally Wood, Luca Carlone, Ali-akbar Agha-mohammadi
TL;DR
Subterranean SLAM must handle degraded sensing, odometry errors, featureless corridors, and spurious loop closures while producing accurate 3D maps. LAMP addresses these conditions with a centralized lidar-based multi-robot system combining scan-to-submap odometry, pose-graph optimization, artifact detection, and ICM outlier rejection. Evaluations show improved odometry and undistorted maps, while the authors identify tuning, scalability, and corridor-degeneracy limitations.
Problem
Subterranean SLAM must produce accurate 3D maps despite poor sensing conditions, inaccurate wheel odometry, lidar drift in featureless corridors, and spurious loop closures.
Method
LAMP combines lidar and vision front-ends with local and centralized pose-graph optimization, using scan-to-submap matching and ICM to reject outlying loop closures.
Results
The system improves odometry over scan-to-scan matching, achieves drift smaller than 1% of traveled distance in the Safety Research mine, and removes loop-closure distortions from the resulting map.
Takeaways & Limitations
LAMP provides robust multi-robot trajectory estimates and constructs point-cloud maps from 3D lidar data in subterranean exploration settings.
Takeaways & Limitations
ICM requires extensive parameter tuning and can fail to balance outlier rejection with retaining inliers, while centralized operation encounters communication bottlenecks with more than four robots.
Abstract
from arXiv · showhide
Simultaneous Localization and Mapping (SLAM) in large-scale, unknown, and complex subterranean environments is a challenging problem. Sensors must operate in off-nominal conditions; uneven and slippery terrains make wheel odometry inaccurate, while long corridors without salient features make exteroceptive sensing ambiguous and prone to drift; finally, spurious loop closures that are frequent in environments with repetitive appearance, such as tunnels and mines, could result in a significant distortion of the entire map. These challenges are in stark contrast with the need to build highly-accurate 3D maps to support a wide variety of applications, ranging from disaster response to the exploration of underground extraterrestrial worlds. This paper reports on the implementation and testing of a lidar-based multi-robot SLAM system developed in the context of the DARPA Subterranean Challenge. We present a system architecture to enhance subterranean operation, including an accurate lidar-based front-end, and a flexible and robust back-end that automatically rejects outlying loop closures. We present an extensive evaluation in large-scale, challenging subterranean environments, including the results obtained in the Tunnel Circuit of the DARPA Subterranean Challenge. Finally, we discuss potential improvements, limitations of the state of the art, and future research directions.
I. INTRODUCTION
Autonomous subterranean exploration supports applications from disaster response to extraterrestrial-world exploration, but degraded sensing and repetitive environments make accurate SLAM difficult. The paper addresses these challenges with a lidar-based multi-robot system evaluated in demanding underground environments.
- Subterranean autonomy is motivated by search and rescue, disaster response, infrastructure monitoring, and exploration of other worlds.
- Accurate 3D mapping is difficult because poor illumination, dust, water, and non-Lambertian surfaces undermine visual SLAM, while uneven terrain degrades wheel odometry.
- Long, featureless corridors make lidar mapping prone to drift, and repetitive tunnel or mine appearances create spurious loop closures that can distort maps.
- Lidar is attractive for complex unstructured environments because it offers a 360° horizontal field of view, high sampling rate, and no dependence on external light.
- The paper contributes a multi-robot architecture with scan-to-scan and scan-to-submap lidar front-end processing, a flexible back-end, and ICM-based rejection of outlying loop closures.
- The system is evaluated extensively in challenging underground mines, including the Tunnel Circuit of the DARPA Subterranean Challenge.
II. LAMP: LARGE-SCALE AUTONOMOUS MAPPING AND POSITIONING
LAMP is a large-scale autonomous mapping and positioning system that uses lidar from one or more robots to estimate trajectories and construct an environment-wide point-cloud map.
- LAMP estimates each robot’s trajectory and a point-cloud map from 3D lidar data, while an RGB-D camera detects and localizes known artifacts.
- The system supports multiple robots connected to a central base station and falls back to single-robot SLAM when communication is unavailable.
A. Overview: Single and Multi-robot Architecture
LAMP combines local lidar and vision front-ends with pose-graph back-ends, then centrally fuses robot estimates and rejects inconsistent inter-robot loop closures.
- A. Overview: Single and Multi-robot Architecture: The local lidar front-end computes odometry and loop closures, while the vision front-end detects and localizes artifacts.
- A. Overview: Single and Multi-robot Architecture: Each robot’s local back-end uses pose-graph optimization to estimate trajectories, artifact positions, and downsampled point clouds attached to poses.
- A. Overview: Single and Multi-robot Architecture: When communication is available, robots send pose estimates, graphs, and local point clouds to a base station for inter-robot loop-closure search and global optimization.
- A. Overview: Single and Multi-robot Architecture: The lidar front-end obtains relative pose measurements between consecutive and non-consecutive poses from raw lidar data.
- A. Overview: Single and Multi-robot Architecture: GICP performs scan-to-scan matching, while scan-to-submap matching is added because compact lidars have limited vertical field of view and resolution.
- A. Overview: Single and Multi-robot Architecture: Loop-closure candidates are found within a 10 m radius and registered with GICP initialized using zero translation and an odometry-derived rotation.
C. Vision Front-end and Artifact Positioning
The vision front-end detects objects of interest and estimates their positions relative to the robot, while the back-end incorporates these detections into the pose graph.
- C. Vision Front-end and Artifact Positioning: The vision front-end detects artifacts and fiducial markers using YOLO and AprilTags, respectively.
- C. Vision Front-end and Artifact Positioning: The back-end uses pose-graph optimization implemented in GTSAM at both each robot and the base station.
- C. Vision Front-end and Artifact Positioning: A key-scan is instantiated after at least 30° of rotation or 1 m of translation, adding the corresponding pose and odometry edge to the local graph.
- C. Vision Front-end and Artifact Positioning: Artifact or fiducial detections add a graph variable for the object position and an edge connecting it to the corresponding robot pose.
2) Base station (Multi-robot) PGO:
The base station merges robot pose graphs and filters inter-robot loop closures before joint pose-graph optimization. Incremental Consistency Maximization combines odometry and pairwise consistency checks to reject closures that violate graph-cycle constraints.
- Base-station graph fusion: The base station merges pose graphs from robots within communication range into a single graph before multi-robot optimization.Robots are assumed to share a common reference frame, which can be established using fiducial markers.
- Outlier rejection: Incremental Consistency Maximization removes loop-closure outliers before pose-graph optimization by combining odometry checks with pairwise consistency tests.The implementation incrementally updates its consistency matrix as new loop closures pass the odometry check.
- Cycle consistency: Loop closures are tested using graph cycles: in the absence of noise, relative pose measurements around a complete cycle should compose to the identity.A loop closure is compared with the chained odometry between the same poses.
- Odometry check: Outlying loop closures produce large accumulated loop errors and are rejected when average rotation error exceeds 0.05 rad or translation error exceeds 0.1 m.The average error is computed over the m edges in the loop.
- Consistent-set selection: After consistency checks, the system finds a mutually consistent set of loop closures by computing the maximum clique of the adjacency matrix.The matrix records which loop-closure pairs are pairwise consistent.
A. Test Environments and Setup
The system was evaluated in six underground tunnels and mines that pose substantial perception and mobility challenges. The Bruceton ground-truth proxy combines artifact constraints, fiducial locations, and sporadic total-station measurements in an optimized pose graph.
- Test environments: The field experiments covered six underground tunnels and mines with challenging perception and mobility conditions.Table I reports explored distance and averaged, rounded tunnel width and height for these environments.
- Robot setup: The robots were Husky-A200 ground vehicles equipped with a VLP-16 lidar, RGB-D camera, onboard computer, and mesh-network radios.The radios connect robots with one another and the base station.
- Base station: The base station was placed outside the mine to receive telemetry, pose graphs, and artifact-localization results from deployed robots.It used an Intel Hades Canyon NUC8i7HVKVA with 32 GB RAM.
- Ground truth: Ground-truth trajectories in Bruceton were approximated by enforcing surveyed artifact locations in each robot’s pose graph and adding sporadic total-station measurements when available.The resulting optimized trajectory was used as the ground-truth proxy shown in Fig. 4.
1) Odometry Evaluation:
Scan-to-submap matching substantially improves lidar odometry over scan-to-scan matching, achieving about 1% drift in nominal conditions. Loop closure detection corrects odometric distortions and can close long loops that standard initialization fails to register.
- Odometry accuracy: Scan-to-submap matching drastically improves odometry accuracy compared with scan-to-scan matching.Scan-to-scan matching quickly drifts because it poorly estimates robot pitch, while scan-to-submap matching performs significantly better.
- Odometry accuracy: 1% drift of the distance traveled is achieved by scan-to-submap odometry in nominal conditions.Drift is smaller than 1% in the Safety Research mine but larger in the Experimental mine because featureless corridors cause GICP to underestimate motion along the tunnel.
- Loop closure detection: Loop closure detection avoids spurious corridors caused by odometry drift in the Eagle and Bruceton Experimental mines.The comparison covers map estimates with and without loop closure detection across three underground mines.
- Loop closure detection: The proposed approach correctly detects and enforces a very long loop in the Beckley Exhibition Coal mine when standard GICP initialization fails.With standard odometric initialization, the loop-closure poses are too far apart for GICP to converge to a reasonable estimate.
3) ICM Evaluation:
ICM rejects inaccurate loop closures that otherwise distort pose-graph maps. LAMP evaluation compares localization with loop closures and outlier rejection across mines and includes a multi-robot mapping experiment.
- ICM evaluation: 0.2 m end-to-end error with ICM is significantly less than the 2.8 m drift without ICM.The comparison uses a robot started and stopped at the same tunnel location; without ICM, inaccurate loop closures visibly distort the map.
- ICM evaluation: ICM removes poor loop closures and produces a final map with no visible distortion.Without ICM, inaccurate loop closures distort the map; with ICM, outlying loop closures are rejected.
- Localization accuracy: Artifact localization error is evaluated with and without loop closures and outlier rejection in the Bruceton Safety Research and Experimental mines.The Tunnel Circuit result has reduced error because total station measurements of robot pose are included in the pose graph.
- Multi-robot mapping: The multi-robot experiment compares fused base-station maps without loop closures, with loop closures but no ICM, and with both loop closures and ICM.The experiment uses two robots at the Bruceton Safety Research mine and evaluates inter-robot loop-closure handling.
IV. DISCUSSION
LAMP demonstrates a centralized multi-robot SLAM system for subterranean mapping, while the discussion identifies important limitations in odometry, loop-closure rejection, and scalability. These limitations motivate future work on sensor fusion, robust global optimization, and more efficient multi-robot communication.
- System outcome: LAMP provides robust trajectories for multiple robots and constructs a point-cloud map from 3D lidar data, with a vision front-end for artifact detection and localization.The system was developed for the DARPA Subterranean Challenge.
- Lidar odometry: Lidar odometry works well in nominal conditions but may underestimate motion in long symmetric corridors lacking detectable geometric features.The scan-matching problem becomes ambiguous in these settings, and GICP can become trapped in local minima.
- Loop-closure detection: The loop-closure strategy finds many opportunities despite odometric drift, but it also produces a large number of outliers.Proposed directions include feature-based global scan matching, vision-based loop closures, and more precise covariance estimation.
- Outlier rejection: ICM offers a pragmatic approach to rejecting outlier loop closures, but requires extensive parameter tuning and may poorly balance outlier rejection against inlier retention.More globally consistent robust solvers could improve reliability and reduce problem-specific tuning, although some current alternatives do not scale well.
- Multi-robot mapping: LAMP’s centralized architecture works well for small robot teams but encounters communication bottlenecks with more than four robots.Future directions include map compression, adaptive data exchange, and compact uncertainty-aware map representations.