Elevator-LIO: A Robust LiDAR-Inertial Odometry in Non-Inertial Frames and Confined Spaces

Paper Code
Elevator-LIO explicitly models elevator-induced non-inertial motion, enabling continuous vertical estimation, stable mapping, and robust LiDAR-inertial localization across floors.

Abstract

This paper presents Elevator-LIO, a LiDAR-inertial odometry framework designed to achieve continuous robot localization during elevator travel, thereby supporting cross-floor robotic tasks. To address the state-estimation problem in non-inertial frames, Elevator-LIO establishes a decoupled state-estimation model that separately models the robot motion relative to the elevator and the elevator motion itself, and embeds it into a mode-dependent iterated error-state Kalman filter framework. This framework degenerates to conventional LIO estimation in ordinary indoor environments, while enabling the propagation and constrained update of elevator-related states in elevator non-inertial environments, thereby achieving continuous and stable localization. An elevator mode manager detects elevator entry and exit events using LiDAR ranging statistics and estimated states, and introduces event-triggered zero-velocity and zero-acceleration updates when the elevator stops to suppress accumulated vertical drift. In addition, this paper adopts an adaptive voxel downsampling strategy to maintain a stable number of effective points under significant environmental scale changes. We conduct extensive experiments on 20 real-world sequences containing 79 elevator rides, including practical challenges such as large-scale spaces, long vertical travel, dynamic pedestrian interference, and mirror reflections. The results show that Elevator-LIO maintains continuous localization accuracy in all sequences, with terminal height error below 1 cm in 16 sequences. In contrast, existing representative localization systems perform poorly on these elevator sequences. Tests on the Hilti 2022/2023 datasets further show that the proposed method remains competitive in standard indoor scenarios.

Closed-Loop Cross-Floor Validation

This sequence validates cross-floor consistency using a closed-loop route: the sensor travels between floors by elevator on one side of the building, then returns via the stairs on the other side. The estimated trajectory closes the loop at the origin, demonstrating that Elevator-LIO recovers the actual vertical displacement and maintains global consistency. Even under aggressive in-cabin motion, the estimator remains stable throughout the sequence.

A closed-loop route combining elevator traversal and stair traversal to validate global cross-floor consistency.

Real-World Elevator Dataset

We evaluate Elevator-LIO on 20 self-collected real-world sequences with 79 elevator rides, covering Office, Dormitory, Campus, and Mall environments. These sequences include large-scale multi-floor mapping, long vertical traversal, handheld in-cabin motion, dynamic pedestrians, and mirrored elevator interiors.

Handheld data acquisition platform
Data are recorded with a compact handheld platform built around a Mid360 LiDAR with integrated IMU, a Jetson Orin Nano, and a synchronized industrial camera.
Overview of Elevator-LIO real-world dataset and terminal vertical error chart
The dataset covers standard, large-scale, long-vertical, and dynamic scenarios, with each sequence visualized by its map, trajectory, and vertical motion profile.

Representative Cross-Floor Mapping Results

Representative real-world sequences show that Elevator-LIO maintains continuous localization across mixed horizontal and vertical motion, producing consistent cross-floor maps in large-scale buildings and remaining stable during long elevator travel.

Large-scale real-world multi-floor mapping results
Representative office and campus sequences demonstrate large-scale cross-floor mapping across complex buildings, including repeated elevator transitions and trajectory revisits.
Long vertical elevator traversal mapping results
Dormitory sequences evaluate long-range vertical consistency, including repeated floor-by-floor traversal and direct return to the starting floor.

Results

Elevator Scenario Results

Scene Sequence Sequence Information Terminal Vertical Error (m) ↓
Elevator Segments
(Detected / Ground Truth)		 Elevator Travel Ratio
(%)		 Trajectory Length
(m)		 w/o ZUPT w/ Reset Full
Office Office1 2/2 29.2 26.4 0.386 0.096 0.002
Office2 2/2 14.9 50.6 1.665 0.223 -0.002
Office3 2/2 3.1 274.5 0.777 0.541 0.001
Office4 2/2 7.3 109.6 -3.074 0.448 -0.002
Office5 2/2 48.0 22.7 -1.877 0.218 0.117
Office6 2/2 6.1 179.3 -0.602 -0.244 0.006
Office7 3/3 44.5 41.8 -1.948 -0.206 -0.136
Office8 3/3 97.4 19.1 0.337 0.212 -0.006
Office9 2/2 4.5 168.4 -0.131 0.160 -0.002
Office10 3/3 43.5 42.8 -1.87699 0.03216 -0.005
Dormitory Dormitory1 3/3 83.2 71.9 -0.003 -0.112 -0.003
Dormitory2 8/8 84.4 92.5 -2.31 -0.033 -0.003
Dormitory3 8/8 85.9 97.5 -0.39 0.017 -0.004
Dormitory4 13/13 88.7 156.6 -0.172 -0.007 -0.007
Campus campus1 2/2 5.7 243.0 0.739 -0.214 -0.005
campus2 6/6 5.3 893.9 -0.781 -0.24586 0.002
campus3 1/1 8.5 148.8 -1.124 -0.662 -0.499
campus4 1/3 4.0 315.1 -0.989 -0.416 0.002
Mall Mall1 6/7 66.4 70.18 -1.749 -0.105 0.002
Mall2 5/5 68.9 67.45 0.738 0.158 0.002

"Elevator Segments" reports the number of detected elevator segments over the ground-truth count; "Elevator Travel Ratio" is the percentage of vertical elevator-travel distance relative to the total trajectory length; "Trajectory Length" is the estimated sequence length. The last three columns report terminal vertical errors for ablated variants: "w/o ZUPT" denotes the variant without the exit-stage zero-velocity/zero-acceleration update; "w/ Reset" clears and rebuilds the map after arriving at a new floor; "Full" denotes the complete Elevator-LIO system and is reported at 1-mm resolution, with absolute errors below 1 mm displayed as 0.001 m. Underlined values indicate the second-best absolute error among the three variants.

Terminal vertical error bar chart for elevator scenarios
Terminal vertical error comparison across the real-world elevator sequences.

Traditional Scenario Results

Dataset Sequence Faster-LIO LIO-SAM Point-LIO VoxelMap FAST-LIO2 Elevator-LIO
(Ours)
Hilti2023 Floor 0 0.016 0.046 0.011 0.017 0.023 0.024
Floor 1 0.028 × 0.025 0.058 0.008 0.017
Floor 2 0.019 0.265 0.087 0.077 0.046 0.107
Stair 0.864 × 0.173 × 0.124 1.262
Underground 1 0.019 0.080 0.041 0.068 0.031 0.013
Underground 2 0.167 × 0.106 0.049 0.125 0.084
Underground 3 0.117 × 0.108 0.075 0.038 0.106
Underground 4 0.022 0.161 0.060 0.036 0.034 0.019
Hilti2022 01 Construction Ground level 0.012 0.078 0.019 1.827 0.015 0.016
02 Construction Multilevel 0.037 1.889 0.059 × 0.026 0.033
03 Construction Stairs × × 3.070 × 0.446 0.331
07 Long Corridor 0.053 × 0.043 0.073 0.071 1.019
11 Lower Gallery 0.810 0.834 1.504 10.337 1.207 0.744
15 Attic to Upper Gallery × × 1.526 × × 2.873
21 Outside Building 0.051 3.061 0.134 23.300 0.248 0.029
06 Constr. Upper Level 2 0.034 0.085 0.030 1.288 0.041 0.047
14 Basement 2 0.166 33.723 0.107 0.622 0.064 0.063

Absolute translational errors are RMSE values in meters. Bold and underlined values denote the best and second-best results respectively; × indicates failure.

Elevator-LIO Overview

Elevator-LIO addresses the reference-frame inconsistency that arises when a robot is carried by a moving elevator. Conventional LiDAR-inertial odometry assumes an inertial world frame, whereas inside an elevator the IMU senses elevator-induced motion and the LiDAR mainly observes cabin-relative geometry.

The system preserves standard LIO behavior in ordinary indoor spaces and switches to a non-inertial elevator mode after cabin entry. Its pipeline integrates sensor buffering, static IMU initialization, adaptive LiDAR downsampling, mode-dependent propagation, IESKF-based LiDAR update, exit-stage correction, and incremental mapping.

Elevator-LIO system overview
System overview of Elevator-LIO. The pipeline combines time-ordered sensor buffering, static IMU initialization, adaptive LiDAR downsampling, elevator mode management, mode-dependent IMU propagation, IESKF-based LiDAR update, exit-stage correction, and incremental ikd-Tree mapping.

Mode-Dependent Propagation

Elevator-LIO decouples the robot motion relative to the elevator from the elevator transport motion. In non-inertial mode, the elevator-related states are propagated by the IMU-driven process model, and the corresponding transition and noise-injection blocks are enabled. In inertial mode, these blocks are disabled, so the propagation reduces to the standard LIO formulation.

Mode-dependent transition matrices
Elevator-related state-transition and noise-injection blocks are activated only in non-inertial mode, while standard LIO propagation is recovered in inertial mode.

Exit Update

At the end of an elevator ride, the cabin becomes stationary. Elevator-LIO uses this event to suppress accumulated vertical drift: it applies zero-velocity and zero-acceleration constraints, re-anchors the estimated elevator displacement into the robot state, and resets elevator-related states to zero.

Event-triggered elevator exit update
The exit update suppresses accumulated vertical drift and restores global consistency when the elevator stops.

Supporting Modules

Elevator entry and exit mode manager

Elevator Mode Manager

Elevator entry is detected by monitoring the drop in the robust maximum LiDAR range as the robot moves from an open space into a closed cabin. Elevator exit is identified from the variance pattern of the estimated vertical elevator velocity.

Adaptive voxel downsampling

Adaptive Downsampling

The voxel size is adjusted online to keep the downsampled point count close to a target value. This preserves geometric details in confined cabins while avoiding excessive computation in open halls.