Multiway registration framework for spatiotemporal environments

Urban Robotics Lab at KAIST, led by Professor Hyun Myung, has developed a multiway registration framework, which robustly aligns LiDAR scans collected across multiple sessions in spatiotemporal environments. This work addresses one of the most critical challenges in SLAM: achieving consistent localization and mapping when structural changes occur over time.

The proposed framework combines three complementary components. CubicFeat is a geometry-aware 3D local feature that leverages hierarchical voxelization and sparse convolution to efficiently extract discriminative descriptors. By focusing on geometrically detailed points, it enables both fast and accurate correspondence matching across scans. Quatro provides a robust global registration solver capable of producing reliable alignments even with very few valid correspondences. Using outlier pruning and quasi-SO(3) estimation, Quatro effectively avoids degeneracy in sparse or partially overlapped point clouds. Finally, Chamelion is a dual-head network for map-to-scan change detection, designed to identify both positive and negative dynamic changes in 3D environments. Within the pipeline, detected changes are removed and the alignment is further refined using generalized iterative closest point (G-ICP), thereby ensuring robustness against structural variations.

Through iterative multiway registration without prior edge information, the pipeline demonstrated reliable graph construction and could tolerate up to 10% outlier scans. The incorporation of change detection allowed the framework to maintain high accuracy even in construction and industrial sites where environmental changes are frequent.

With this technology, the KAIST team achieved first place overall at the Nothing Stands Still (NSS) Challenge 2025, held at the IEEE International Conference on Robotics and Automation (ICRA) in Atlanta, USA. The KAIST team achieved mapping error of 0.28m over 15,000 m² area while the second place team only achieved 7.98m error. Competing against leading research institutions worldwide, the team’s framework showcased both academic innovation and practical applicability to autonomous driving, construction robotics, legged platforms, and aerial systems.

Professor Myung noted, “Our multiway registration framework shows how SLAM can adapt to the reality of changing environments. Winning the NSS Challenge validates our approach and highlights KAIST’s commitment to advancing the frontier of robotics.”