Source-linked AI summary
LAMP 2.0: A Robust Multi-Robot SLAM System for Operation in Challenging Large-Scale Underground Environments
Yun Chang, Kamak Ebadi, Christopher E. Denniston, Muhammad Fadhil Ginting, Antoni Rosinol, Andrzej Reinke, Matteo Palieri, Jingnan Shi, Arghya Chatterjee, Benjamin Morrell, Ali-akbar Agha-mohammadi, Luca Carlone
TL;DR
LAMP 2.0 addresses robust localization and mapping for heterogeneous robot teams in challenging, large-scale underground environments. It integrates adaptable per-robot interfaces, scalable loop-closure detection, and outlier-robust pose-graph optimization, achieving below-2 m trajectory errors on trajectories up to 2.2 km while releasing benchmark datasets.
Problem
High-precision multi-robot SLAM is needed for underground search and rescue, where darkness, dust, rugged terrain, and ambiguous scenes degrade perception.
Method
LAMP 2.0 combines adaptable single-robot interfaces, scalable intra- and inter-robot loop-closure detection, and outlier-robust centralized pose-graph optimization.
Results
Below 2 m average trajectory errors were achieved for trajectories up to 2.2 km in end-to-end tests across four datasets.
Takeaways & Limitations
The system provides field-tested cooperative localization and mapping for heterogeneous multi-robot teams in challenging underground environments.
Abstract
from arXiv · showhide
Search and rescue with a team of heterogeneous mobile robots in unknown and large-scale underground environments requires high-precision localization and mapping. This crucial requirement is faced with many challenges in complex and perceptually-degraded subterranean environments, as the onboard perception system is required to operate in off-nominal conditions (poor visibility due to darkness and dust, rugged and muddy terrain, and the presence of self-similar and ambiguous scenes). In a disaster response scenario and in the absence of prior information about the environment, robots must rely on noisy sensor data and perform Simultaneous Localization and Mapping (SLAM) to build a 3D map of the environment and localize themselves and potential survivors. To that end, this paper reports on a multi-robot SLAM system developed by team CoSTAR in the context of the DARPA Subterranean Challenge. We extend our previous work, LAMP, by incorporating a single-robot front-end interface that is adaptable to different odometry sources and lidar configurations, a scalable multi-robot front-end to support inter- and intra-robot loop closure detection for large scale environments and multi-robot teams, and a robust back-end equipped with an outlier-resilient pose graph optimization based on Graduated Non-Convexity. We provide a detailed ablation study on the multi-robot front-end and back-end, and assess the overall system performance in challenging real-world datasets collected across mines, power plants, and caves in the United States. We also release our multi-robot back-end datasets (and the corresponding ground truth), which can serve as challenging benchmarks for large-scale underground SLAM.
I. INTRODUCTION
LAMP 2.0 addresses robust cooperative localization and mapping for heterogeneous robot teams operating in large, perceptually degraded subterranean environments. It extends LAMP with scalable loop-closure detection, outlier-robust optimization, and released benchmark data.
- Motivation: Subterranean search and rescue requires high-precision localization and mapping despite complex underground conditions.The DARPA Subterranean Challenge targeted robotic exploration and mapping to support localization of elements of interest such as survivors.
- Related work: Earlier subterranean systems used lidar, vision, thermal, and inertial sensing for localization and mapping in confined environments.Prior work included photogrammetry with parallel vision- and lidar-based SLAM and loosely coupled multimodal fusion.
- Prior system: LAMP previously combined lidar scan matching with pose-graph optimization but struggled during larger-scale, long-duration operation under perceptual aliasing.The limitation concerns unknown subterranean environments where scenes can appear self-similar.
- Contributions: LAMP 2.0 contributes a computationally efficient, outlier-resilient centralized multi-robot SLAM system adaptable to different odometry sources.Its improved modules include scalable loop-closure detection and pose-graph optimization based on Graduated Non-Convexity.
- Contributions: The paper releases LAMP 2.0 and multi-robot subterranean datasets with ground-truth trajectories and maps for evaluating localization and mapping methods.The datasets cover caves, mines, and abandoned power plants.
II. LAMP 2.0
LAMP 2.0 combines per-robot interfaces, centralized multi-robot loop-closure detection, and robust back-end optimization for localization and map generation.
- Architecture: Each robot sends pose-graph segments and keyed scans from odometry and point-cloud streams to the centralized base station.The centralized front end receives local odometry and maps for multi-robot processing.
- Architecture: LAMP 2.0 uses a single-robot front end, a centralized multi-robot front end, and a multi-robot back end.The architecture is designed for heterogeneous robots and underground operation.
- Architecture: The multi-robot front end detects intra-robot and inter-robot loop closures, while the back end optimizes trajectories and generates maps.The back end uses odometry and loop closures from all robots.
A. Initial Pose Calibration
Initial pose calibration establishes a common reference frame for all robots before subterranean operation.
- Calibration: Before operation, the robots’ sensors are assumed to have intrinsic and extrinsic calibration.Initial pose calibration then aligns the team in a shared reference frame.
- Calibration: The calibration procedure places three reflective fiducial plates with known 3D coordinates on an entrance gate.The robots scan the gate while positioned in front of it.
- Calibration: Alternative methods can establish a common reference frame without initial pose calibration.The gate-based method was used for compatibility with the SubT Challenge.
B. Adaptable Single-robot Front-end Interface
The adaptable single-robot front end lets LAMP 2.0 process heterogeneous odometry and lidar inputs into compact pose-graph and keyed-scan representations.
- Adaptability: The interface supports different odometry sources and lidar configurations, including robots using LOCUS or Hovermap.A single experiment can include robots with different lidar numbers and configurations.
- Point-cloud processing: HeRO corrects motion distortion in lidar scans before the system processes the point clouds.It computes a local state estimate and uses relative motion during scan acquisition for correction.
- Point-cloud processing: The front end merges calibrated lidar clouds and applies adaptive voxelization to remove redundancy and noise while retaining a constant point count.This supports inputs from different lidar configurations and varying environment geometry.
- Pose-graph generation: LAMP uses sparse pose graphs, adding key nodes after motion thresholds such as 2 m translation or 30 degrees rotation.Each key node stores a keyed scan, and graphs can be batched after communication is re-established.
C. Scalable Multi-robot Front-end
The multi-robot front-end scales loop-closure detection by generating, prioritizing, and computing candidates while avoiding ambiguous or geometrically degenerate matches.
- Loop closures reduce trajectory drift and merge overlapping robot maps into a geometrically consistent global map.
- Loop closure generation: Adaptive distance thresholds generate loop-closure candidates according to relative traversal for single robots and absolute traversal for multi-robot teams.The threshold uses dmax = α|ncurr − ncand| or dmax = αncurr, with α set to 0.2 m in experiments.
- Loop closure prioritization: Prioritization orders candidates by their expected contribution to improving the trajectory estimate.
- Loop closure prioritization: The system addresses perceptual aliasing by prioritizing feature-rich areas and graphically valuable closures while using RSSI-based known locations when available.Feature richness is estimated from eigenvalues of the point-to-plane ICP information matrix, while a GNN predicts pose-graph optimization outcomes.
- Loop closure computation: Each queued candidate receives an initial alignment with TEASER++ or SAC-IA, followed by GICP refinement and filtering.
D. Robust Multi-robot Back-end
The centralized back-end combines all robots’ odometry with intra- and inter-robot loop closures to estimate globally consistent trajectories and maps while rejecting erroneous constraints.
- The centralized back-end performs pose graph optimization using odometry and loop closures from all robots to generate globally consistent, drift-free maps.
- Outlier-robust optimization computes a consistent trajectory estimate for the entire heterogeneous robot team.
- ICM checks loop-closure consistency before graph insertion, while GNC with Levenberg–Marquardt jointly estimates trajectories and inlier/outlier decisions.
III. EXPERIMENTS
The experiments evaluate LAMP 2.0’s architecture across heterogeneous robots and sensor configurations using real-world underground datasets, with computation performed on specified workstation and laptop platforms.
- The evaluation covers hardware and system performance, component ablations, open-source datasets, and all three SubT challenge events.
- Robots: Husky and Spot robots use varying lidar configurations and either LOCUS or Hovermap odometry, exposing the front-end to heterogeneous inputs.
- Base station: The SubT experiments used a 64-core AMD Ryzen Threadripper workstation, while paper experiments ran on a 12-core Intel i7 laptop.
B. Datasets
Four Team CoSTAR datasets span coal mines, abandoned power plants, DARPA-built courses, and limestone mines, with surveyed-map ground truth and released pose-graph data for benchmarking.
- The four datasets cover diverse underground settings and team scales, with combined traversals ranging from 1.2 km to 6 km.They include Tunnel, Urban, Finals, and Kentucky Underground datasets collected with Husky and Spot teams.
- Ground-truth trajectories derive from surveyed global maps using scan-to-map localization, and ground-truth and false loop-closure pairs stress robustness.
- The released pose graphs and keyed scans are intended to support research on multi-robot loop-closure detection and robust pose graph optimization.
C. Component Evaluation
Component evaluations show that LAMP 2.0 improves loop-closure quality through two-stage initialization and obtains more robust trajectory estimates with GNC-based outlier rejection.
- Evaluation scope: The component evaluation examines loop-closure detection, outlier rejection, and the quantity of loop closures surviving frontend and backend stages.The study compares the new modules with the previous LAMP configuration.
- Two-stage loop closure computation: The two-stage loop-closure pipeline evaluates odometric, TEASER++, and SAC-IA initializations before ICP refinement using recall, false-positive rate, and pose-error metrics.Correct loop closures must pass SAC and ICP error thresholds; pose errors are computed against the ground-truth trajectory.
- Two-stage loop closure computation: TEASER++ and SAC-IA reduce false-positive rates and mean translation and rotation errors by filtering poor alignments with insufficient inliers.These methods improve the accuracy of accepted loop closures, although recall varies with environment width.
- Two-stage loop closure computation: Recall depends on environment geometry: odometric initialization performs better in narrower environments, while TEASER++ and SAC-IA perform better in wider environments.The wider-environment examples include Urban and KU.
- Outlier-robust pose graph optimization: GNC generally produces more robust and accurate trajectory estimates than ICM, especially on the longer Tunnel and Kentucky Underground datasets.Both rejection methods improve estimates relative to no loop closures or no outlier rejection, but ICM misses some outliers or becomes too conservative.
D. System Evaluation
End-to-end playback evaluates LAMP 2.0 against LAMP 1.0 and a version without inter-robot loop closures across four datasets. The system achieves trajectory errors below 2 m for trajectories up to 2.2 km and map errors below 4 m in challenging large-scale environments.
- Trajectory and map evaluation: LAMP 2.0 achieves average trajectory errors below 2 m with trajectory lengths of up to 2.2 km across the evaluated datasets.The comparison includes LAMP 1.0 and LAMP 2.0 without inter-robot loop closures, using consistent odometry inputs and real-time playback.
- Trajectory and map evaluation: LAMP 2.0 achieves map errors below 4 m across four large-scale underground environments.The maps are evaluated against ground-truth maps using cloud-to-cloud error.
IV. CONCLUSION
LAMP 2.0 is a centralized multi-robot SLAM system that estimates multiple robot trajectories and constructs a 3D lidar point-cloud map. Its experiments show strong performance with up to four robots, while centralized operation remains a scalability boundary.
- LAMP 2.0 estimates multiple robot trajectories and constructs a point-cloud map using 3D lidar data.
- Its single-robot front end accepts different odometry inputs and lidar configurations, while the back end combines prioritized loop closures with outlier-robust pose graph optimization.
- The system performed well in experiments with up to four robots.
- LAMP 2.0’s centralized architecture may not scale to large robot teams.
- Extending LAMP to a fully distributed system with intermittent inter-robot communications is identified as future work.