CDC-SLAM: Switchable Centralized and Distributed Collaborative LiDAR SLAM Framework for Robotic Swarms
Xiangnan Liu, Xiang Huo, Haifei Zhu, Xuefeng Zhou, Yisheng Guan, Hong Zhang, Weinan Chen∗
AI summary
Problem
Existing multi-robot SLAM frameworks force a trade-off between the robustness of centralized architectures and the scalability of distributed ones, leaving a gap in reliably handling diverse, large-scale swarm scenarios.
Approach
The system monitors inter-robot distances to trigger seamless mode switching, while a dynamic pose fusion mechanism corrects frontend drift without causing trajectory jumps.
Key results
- Adaptive distance-triggered switching between centralized and distributed optimization modes
- Dynamic pose fusion method that prevents trajectory discontinuities during mode transitions
- Superior localization accuracy and mapping consistency validated on KITTI and S3E datasets
- Open-source release of code and processed multi-robot datasets
Why it matters
Provides a practical, scalable solution for large-scale robotic swarm navigation in GPS-denied environments, advancing multi-robot autonomy research and deployment.
Abstract
In GPS-denied environments, robotic swarms must concurrently accomplish collaborative tasks and au- tonomous localization, a process that remains highly chal- lenging. Existing mainstream collaborative frameworks are generally categorized into centralized and distributed (or decen- tralized) approaches: centralized methods often exhibit limited robustness, while distributed or decentralized methods typically encounter accuracy constraints. The intrinsic limitations of these two framework types render a single architecture inade- quate for effectively addressing complex and diverse real-world scenarios. To address this challenge, we introduce CDC-SLAM, an adaptive centralized–distributed collaborative SLAM frame- work that supports flexible dual-architecture switching, thereby enhancing scenario adaptability and enabling efficient LiDAR- inertial collaborative state estimation. The system adaptively selects optimization strategies according to real-time inter-robot distances: when the swarm is relatively close, a central node performs global centralized optimization and disseminates the results; when dispersed, the system switches to distributed optimization, exchanging only essential data to alleviate the computational burden on individual robots.Meanwhile, back- end data sharing and outlier rejection mechanisms preserve the consistency of the global map during switching. Extensive evaluations on public datasets demonstrate that the proposed CDC-SLAM system delivers improved localization accuracy and mapping performance. IndexTerms—Distributed, Centralized, Multi-robot SLAM.