Chain-SLAM: Globally Consistent Backend for Multi-Session LiDAR SLAM via Chained Loop Closure

Abstract

Staying geometrically consistent over long spatial and temporal horizons is a fundamental challenge in large-scale LiDAR SLAM, especially when maps from many separate sessions must be merged. Chain-SLAM is a LiDAR SLAM backend for online multi-session map alignment and reuse at scale. A chained loop closure mechanism propagates each reliable short-horizon closure across inter-session keyframes through an adjacency graph, yielding long-horizon consistency; sessions are brought into a common frame by GNSS-proximity place recognition and then jointly optimized in a unified factor graph. The result is inter- and intra-session consistency without dynamic object removal, and cross-platform robustness with minimal tuning.

Overview Video

Chained Loop Closure

Chained closure detection and verification pipeline — A new direct loop closure seeds a BFS over the pose graph; ICP verifies each candidate chained closure (keeping those with residual *e_ij ≤ τ*), and the verified constraints become new edges in the cross-session adjacency graph.

Comparison: With vs. Without Chained Closure

The videos below show five sessions accumulating over an intersection that each of them revisits from a different direction, with loop closure constraints drawn in magenta. With direct closures alone the session trajectories stay fanned apart through the turn; chaining adds the constraints that pull them onto the same geometry.

w/o chained closure

w/ chained closure

Multi-Session Map Alignment

The current session starts at the origin of its own frame, inconsistent with the map frame of the loaded sessions. A RANSAC-estimated transform aligns it to the loaded map, which re-enables loop closure detection against the earlier sessions along the transformed trajectory.

Before merging

Current session trajectory in its own coordinate frame, misaligned with the loaded map

After merging

Current session trajectory transformed into the loaded map frame

Results

Trajectory evaluation across sessions — Per-session trajectories and alignment error against ground truth, on the multi-session MARS and NCLT datasets.

Trajectory Accuracy (3D APE)

Absolute pose error over full 6-DoF trajectories, with first and second places highlighted; lower is better. Chained loop closure gives the lowest translation error on both datasets.

Dataset	Method	T RMSE (m)	R RMSE (deg)
MARS	Fast-LIO2	52.616	4.708
	LAMM	43.688	10.285
	Ours unchained	13.511	3.729
	Ours chained	9.640	3.448
NCLT	Fast-LIO2	2.279	2.618
	LAMM	33.547	11.549
	Ours unchained	1.473	4.876
	Ours chained	1.197	4.647

Runtime and Memory

Measured on the final session, with all four previous maps loaded and jointly optimized. Frontend cost stays at the Fast-LIO2 level because optimization runs asynchronously in the backend: roughly 2 Hz on the 12-hour, 128-channel MARS dataset and 13 Hz on the 8-hour, 32-channel NCLT dataset.

Dataset	Method	Frontend (ms)	Backend (ms)	FPS (Hz)	RAM (MB)
MARS	Fast-LIO2	22.28	–	44.88	1065.56
	LAMM	22.28	21.25	22.97	16821.90
	Ours unchained	29.88	402.33	2.31	12571.12
	Ours chained	29.39	447.02	2.10	13668.46
NCLT	Fast-LIO2	7.34	–	136.17	928.77
	LAMM	7.34	11.31	53.61	5917.86
	Ours unchained	7.79	67.50	13.28	5682.82
	Ours chained	7.82	70.13	12.83	5816.96

Full tables and ablations are in the paper.

BibTeX

@inproceedings{li2026chainslam,
  title     = {Chain-SLAM: Globally Consistent Backend for Multi-Session LiDAR SLAM via Chained Loop Closure},
  author    = {Li, Zhiheng and Liu, Xinhao and Zhang, Juexiao and Liang, Yongqing and Feng, Chen},
  booktitle = {IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS)},
  year      = {2026}
}