Abstract
Collaborative robotic tasks require a shared understanding of the environment and robot locations. Distributed Simultaneous Localization and Mapping (SLAM) provides a solution without relying on external positioning like GPS. However, existing systems often struggle with perception outliers, leading to conservative parameters that sacrifice accuracy. This paper presents DOOR-SLAM, a fully distributed system featuring a robust outlier rejection mechanism. By utilizing peer-to-peer communication and a distributed pose graph optimizer, DOOR-SLAM maintains high accuracy in challenging environments, including GPS-denied subterranean areas, while significantly reducing communication bandwidth.
Introduction: The Challenge of Multi-Robot Mapping
For a team of robots to coordinate effectively, they must build and maintain a consistent shared map. Distributed SLAM is the standard approach for this task, yet it faces two primary hurdles:
- Communication Constraints: Robots often operate with limited bandwidth and intermittent connectivity, precluding the use of a central server.
- Perceptual Aliasing: In repetitive environments (e.g., identical hallways or open fields), robots may incorrectly identify different locations as the same. These "false loop closures" can lead to catastrophic map distortions.
Traditional systems are often overly conservative, rejecting many valid loop closures to avoid outliers. DOOR-SLAM addresses this by being "brave but smart"—it accepts more candidates but employs a decentralized filter to purge incorrect data.
Methodology
DOOR-SLAM is built on a decentralized framework using the Buzz programming language, consisting of two primary modules.
Distributed Loop Closure Detection
To minimize bandwidth, robots do not exchange raw sensor data. Instead, they share compact NetVLAD descriptors—global signatures of images. When a match is found, robots perform geometric verification using RANSAC and PnP to estimate relative poses. This approach reduces required bandwidth by a factor of 10 compared to raw image exchange.
Distributed Robust Pose Graph Optimization (PGO)
The core innovation of DOOR-SLAM is its outlier rejection mechanism based on Pairwise Consistent Measurement (PCM) set maximization.
- Consistency Logic: The system checks if pairs of loop closures are mathematically consistent with the robots' internal odometry.
- Maximum Clique Algorithm: Consistency checks are treated as a graph problem. The system identifies the "maximum clique"—the largest set of mutually consistent measurements.
- Optimization: After purging outliers, robots use a Distributed Gauss-Seidel approach to reach a globally consistent trajectory estimate without a central coordinator.
Experimental Evaluation
The system was validated across diverse scenarios, including the ARGoS simulator for large-scale drone testing and the KITTI00 automotive benchmark. Real-world field experiments involved quadcopters flying over a football field—an environment prone to perceptual aliasing due to uniform grass—and Husky UGVs in a DARPA Subterranean Challenge coal mine, proving the system's sensor-agnostic nature.
Key Results and Insights
DOOR-SLAM demonstrates that effective outlier rejection allows for more aggressive place recognition, finding up to 3x more valid loop closures than conservative systems.
- Accuracy: On the KITTI00 dataset, DOOR-SLAM reduced the Average Translation Error (ATE) from 86.85 meters (with outliers) to just 8.00 meters.
- Efficiency: The system runs online on NVIDIA Jetson TX2 onboard computers, making it suitable for real-time edge deployment.
- Robustness: In the DARPA SubT coal mine, the system successfully corrected map deformations caused by incorrect LiDAR matches in tunnel segments.
Conclusion and Future Work
DOOR-SLAM provides a robust, bandwidth-efficient framework for decentralized multi-robot SLAM. While the "maximum clique" problem is NP-hard, the current implementation handles hundreds of loop closures effectively. Future research will focus on scaling to tens of thousands of candidates, improving covariance estimation for sensor noise, and increasing resilience to network packet loss.
Full source code is available at: https://github.com/MISTLab/DOOR-SLAM.git
Summary updated