💻 computer science

Elevator-LIO: Robust LiDAR-Inertial Odometry for Multi-Floor Navigation under Elevator-Induced Non-Inertial Motion

Elevator-LIO is a robust LiDAR-inertial odometry framework that ensures continuous and accurate robot localization during elevator travel by employing a decoupled state-estimation model within a mode-dependent Kalman filter, effectively handling non-inertial motion and suppressing vertical drift to support cross-floor navigation.

Original authors: Yifan Zhang, Yudong Huang, Yuchong Zhang, Changze Li, Haoran Liu, Ming Yang, Tong Qin

Published 2026-05-26

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Yifan Zhang, Yudong Huang, Yuchong Zhang, Changze Li, Haoran Liu, Ming Yang, Tong Qin

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are trying to navigate a robot through a building. Usually, this is like walking through a park: you can see the trees, the paths, and the sky, so you know exactly where you are. But what happens when your robot steps into an elevator?

Suddenly, the world changes. The robot is inside a metal box that is moving up or down at high speed. This creates a tricky problem for the robot's "brain" (its navigation system).

The Problem: The "Moving Box" Confusion

Most robot navigation systems work on a simple rule: "If the sensors feel acceleration, the robot is moving."

In a hallway: If the robot feels a push forward, it knows it is walking forward.
In an elevator: If the elevator shoots up, the robot feels a heavy push. But the robot isn't moving up; the box is.

Old navigation systems get confused here. They think, "Whoa, I'm flying!" and start drawing a map where the robot is floating into the sky. By the time the elevator stops, the robot thinks it's on the 50th floor when it's actually on the 2nd. This is called drift, and it causes the robot to get lost.

The Solution: Elevator-LIO

The paper introduces Elevator-LIO, a new navigation system designed specifically to solve this "moving box" problem. Think of it as giving the robot a special two-part brain.

1. The "Decoupled" Brain (Separating the Box from the Robot)

Instead of trying to figure out everything at once, Elevator-LIO splits the problem into two separate stories:

Story A: How the robot is moving inside the elevator (walking around, turning).
Story B: How the elevator itself is moving up or down.

It's like being on a train. You can walk down the aisle (Story A) while the train speeds through a tunnel (Story B). Elevator-LIO tracks both stories separately so it doesn't confuse your walking with the train's speed. It uses the robot's accelerometer (which feels the shake) and its laser scanner (which sees the elevator walls) to keep these two stories straight.

2. The "Pit Stop" Correction (Zeroing the Drift)

Even with a smart brain, small errors can add up over a long ride. Imagine trying to guess how far you've walked in the dark; you might be off by a few inches after a minute.

Elevator-LIO has a clever trick for when the elevator stops. When the doors open, the system knows the elevator is sitting still. It says, "Okay, the elevator isn't moving right now. Let's reset our guess to zero." It uses this moment of stillness to correct all the tiny errors that built up during the ride, ensuring the robot knows exactly which floor it's on.

3. The "Smart Camera" (Adaptive Downsampling)

Elevators are tiny, while building lobbies are huge.

In a lobby: A standard camera sees too many details, overwhelming the computer.
In an elevator: A standard camera sees too few details, making it hard to tell where you are.

Elevator-LIO uses a "smart camera" that adjusts its focus automatically. When the robot enters a small elevator, it zooms in (keeps more details). When it leaves into a big lobby, it zooms out (ignores extra details). This keeps the computer running smoothly no matter how big or small the room is.

The Results: Does it Work?

The researchers tested this system on 20 real-world scenarios, including offices, dorms, and shopping malls, with 79 elevator rides.

The Competition: They compared it to other top navigation systems. The old systems failed miserably, often thinking the robot was floating or getting completely lost inside the elevator.
The Winner: Elevator-LIO kept the robot on track the entire time. In most tests, it was accurate to within 1 centimeter (less than half an inch) of the correct floor height.

Summary

Elevator-LIO is like giving a robot a special pair of glasses that can tell the difference between "I am walking" and "The floor I'm standing on is moving." By separating these two movements and taking a quick "reality check" whenever the elevator stops, it allows robots to navigate multi-story buildings without getting confused or lost.

Technical Summary: Elevator-LIO

Problem Statement

Autonomous mobile robots (AMRs) require robust multi-floor navigation capabilities, yet elevator travel presents a critical failure point for current localization systems. Standard LiDAR-Inertial Odometry (LIO) assumes the sensor platform operates within an inertial reference frame. This assumption breaks down inside a moving elevator, where the platform enters a non-inertial environment. In this scenario, IMU measurements are contaminated by the elevator's vertical transport motion, while LiDAR observations are restricted to the relative geometry of the cabin interior. Conventional LIO pipelines incorrectly attribute elevator-induced acceleration to the robot's ego-motion, leading to severe vertical drift, map inconsistency, or system divergence. Furthermore, the abrupt transition between open floor spaces and confined elevator cabins causes drastic changes in scene scale, destabilizing fixed-resolution point-cloud downsampling strategies.

Methodology

The paper proposes Elevator-LIO, a LiDAR-inertial odometry framework designed to maintain continuous localization during elevator traversal. The system is built upon three core technical pillars:

1. Decoupled Non-Inertial State Estimation

Elevator-LIO introduces a unified state-estimation framework that explicitly separates the robot's motion relative to the elevator from the elevator's absolute motion in the world frame.

Augmented State: The system state is augmented to include elevator transport states ( $p^W_{Ez}, v^W_{Ez}, a^W_{Ez}$ ) representing vertical displacement, velocity, and acceleration in the world frame, alongside the standard robot-relative states ( $p^E_I, q^E_I, v^E_I$ ).
Kinematic Modeling: The framework derives decoupled motion equations for the non-inertial elevator frame. It models the elevator acceleration ( $a^W_{Ez}$ ) as a random walk state rather than a robot-relative acceleration, allowing the smoother component of motion to be estimated online.
Mode-Dependent Filtering: The system utilizes a mode-dependent Iterated Error-State Kalman Filter (IESKF). In ordinary indoor environments, it degenerates to a standard LIO pipeline. Upon detecting elevator entry, it switches to the non-inertial mode, propagating both robot and elevator states. The LiDAR update constrains only the robot-relative state, while the elevator transport state is propagated via the IMU model until external constraints become available.

2. Drift Suppression via Event-Triggered Updates

To address the unobservability of absolute vertical motion during travel, the system employs specific correction mechanisms:

Elevator Mode Manager: A lightweight detector identifies entry and exit events. Entry is triggered by a sharp drop in LiDAR depth metrics (indicating spatial confinement), while exit is detected via the variance pattern of the estimated vertical velocity.
Zero-State Updates (ZUPT/ZAA): When the elevator stops (detected at the exit phase), the system applies high-confidence zero-velocity and zero-acceleration constraints to the elevator transport states. This suppresses accumulated vertical drift.
State Re-anchoring: Upon exiting the elevator, the estimated vertical displacement is absorbed into the robot's global position state, and the auxiliary elevator states are reset, ensuring a seamless transition back to standard inertial LIO.

3. Geometry-Aware Adaptive Downsampling

To handle the drastic scale changes between open floors and confined cabins, the front-end employs a feedback-driven adaptive downsampling strategy. Instead of a fixed voxel resolution, the system dynamically adjusts the voxel size to maintain a target number of effective points. This ensures stable geometric constraints and computational efficiency regardless of the environment's spatial density.

Key Contributions

Decoupled Motion Formulation: The derivation of decoupled motion equations for elevator-induced non-inertial frames, enabling the simultaneous estimation of robot-relative pose and absolute elevator transport states.
Drift-Suppression Mechanism: A tailored correction strategy combining zero-state updates at elevator stops with a lightweight, onboard sensor-based elevator state detector, improving robustness against uncertainty accumulation.
Adaptive Front-End: A feedback-driven adaptive downsampling strategy that maintains consistent point-cloud density across abrupt scene-scale transitions.
Comprehensive Validation: Extensive experiments on 20 real-world sequences (79 elevator rides) covering diverse challenges (large-scale spaces, long vertical travel, dynamic interference, mirror reflections) and public benchmarks (Hilti 2022/2023).

Experimental Results

Real-World Elevator Performance: In 20 real-world sequences, Elevator-LIO maintained continuous localization accuracy in all cases. The terminal height error was below 1 cm in 17 sequences. In contrast, representative baseline methods (FAST-LIO2, LIO-SAM, Point-LIO, VoxelMap) suffered from Type-I failures (drift) or Type-II failures (divergence) in all tested elevator sequences.
Standard Environment Performance: On the Hilti 2022 and 2023 datasets (standard inertial environments), Elevator-LIO remained competitive, achieving the best or second-best performance on several sequences, demonstrating that the proposed reformulation does not degrade performance in ordinary scenarios.
Robustness: Simulation experiments confirmed that the system remains stable and bounded even under controlled model deviations, such as time-varying gravity drift and cabin inclination errors.

Significance and Claims

The paper claims that Elevator-LIO addresses a key bottleneck in multi-floor robotic autonomy: reliable state estimation during the non-inertial motion of elevator travel. By explicitly modeling the coupling between elevator transport and robot motion, the system overcomes the physical inconsistencies that cause conventional LIO to fail. The authors emphasize that their approach enables continuous, divergence-free localization without relying on external infrastructure (e.g., building-side sensors or floor plans). While the current implementation focuses on vertical translation in elevators, the framework provides a foundation for future research into localization on other moving non-inertial carriers (e.g., subways, ships) and the integration of multimodal state management for more complex environments.