EliOT: A Purely Data-Driven, Robust, and Flexible LiDAR Odometry system

Researchers develop a data-driven LiDAR odometry system that avoids computationally intensive geometric methods while being robust and scalable

LiDAR odometry is a vital remote-sensing technology, used to estimate motion of robots and vehicles. Existing LiDAR methods use either memory-intensive geometry-based approaches or learning-based methods that rely on lossy 3D-to-2D projections. In a new study, researchers developed EliOT: end-to-end LiDAR odometry transformers, which estimates motion directly from data. It leverages real-world, simulation, and digital twin data, processes raw 3D information, making it robust and flexible across environments, offering a promising solution for urban navigation.

Image title: Working of EliOT (End-to-End LiDAR odometry with transformers) for robust 3D LiDAR odometry

Image caption: EliOT uses self-attention-based flow embeddings and 3D-transformer networks to estimate motion directly from raw 3D data, offering a robust and scalable LiDAR odometry-based navigation solution for urban environments.

Image credit: Dr. Daegyu Lee from the Electronics and Telecommunications Research Institute

License type: Original content

Usage restrictions: Cannot be reused without permission

LiDAR odometry (LO) is a vital technology used to estimate the motion of robots or vehicles, especially in environments where global navigation satellite systems (GNSS) may be unreliable. LiDAR, short for Light Detection and Ranging, is used to measure distances by sending out laser pulses and measuring the time required for them to bounce back from objects. It produces point clouds, which are collections of millions of three-dimensional (3D) points that provide a virtual representation of objects like buildings and trees. LO uses two consecutive point cloud scans to estimate motion.

The goal of LO is to minimize translational and rotational errors when estimating robot motion. Existing LO algorithms rely on either geometry-based or data-driven, learning-based methods for this purpose. While geometry-based methods have good accuracy, they are memory intensive and require heavy tuning and adjustments for dynamic environments. On the other hand, learning-based methods are more efficient and work well in dynamic conditions. However, they rely on 3D-to-2D projections which cannot fully reflect 3D characteristics and often result in the loss of critical information.

To address these issues, a research team led by Dr. Daegyu Lee from the Electronics and Telecommunications Research Institute in South Korea, developed EliOT: end-to-end LiDAR odometry with Transformers. “Unlike computationally expensive geometry-based approaches, EliOT learns directly from data while also avoiding 3D-to-2D projections, making it more efficient and flexible than traditional LO algorithms,” explains Dr. Lee. “It leverages real-world, simulated, and digital twin-data, and gets better and smarter as more data is added.” Their findings were published in the ETRI Journal in 2025.

First, EliOT extracts geometric features from consecutive point cloud data using a learning-based approach. This generates key points and their corresponding subsampled set of abstracted features. This data is then fed into an implicitly represented flow embedding (IRFE) module. IRFE, a key innovation of this study, uses self-attention-based methods to integrate both local and global spatial contexts and infers motion patterns directly from data without relying on memory-intensive geometry-based methods. This leads to more robust and flexible feature representation, even in dynamically changing environments.

Finally, the output from the IRFE module is fed into a 3D transformer module, which processes raw 3D data to estimate the robot’s pose. This eliminates the need for lossy 3D-2D projections. For training EliOT, the researchers leveraged diverse datasets, including real-world data from the well-known KITTI odometry dataset, simulated data from CARLA (an autonomous driving simulator), and data from a digital twin of the ETRI campus.

As a result, EliOT demonstrated competitive performance in comparison to both classical geometry-based and recent learning-based LO methods. Notably, in evaluations using the KITTI dataset, the model demonstrated higher translation and rotation accuracy compared to previous learning-based methods. In CARLA simulations and the ETRI digital twin environment, the model demonstrated high accuracy in measuring relative translational motion and rotation. This indicates that EliOT can be effectively applied in sim-to-real-to-sim applications and provide high-quality data.

“Our learning-based method will prove valuable for applications like autonomous cars, delivery drones, and disaster-response drones, navigating dense urban areas where GPS is unreliable,” says Dr. Lee. “Importantly, by using this model, vehicles and robots won’t need exhaustive retraining in every new city or environment. Simulation and digital twin exposure will ensure robustness across diverse conditions.”

Overall, EliOT represents a new data-driven paradigm in LO, offering robust and scalable solutions for urban navigation challenges.

Reference

Title of original paper:

ELiOT:End-to-End LiDAR Odometry with Transformers Harnessing Real-World, Simulated, and Digital Twin

Journal: ETRI Journal

DOI: https://doi.org/10.4218/etrij.2025-0011

About the Electronics and Telecommunications Research Institute (ETRI)

Established in 1976, the Electronics and Telecommunications Research Institute (ETRI) is a non-profit government-funded research institute and is one of the leading research institutes in the wireless communications domain. It has more than 2500 patents filed. Equipped with state-of-the-art labs, this institute strives for social and economic development through technology research.

About Dr. Daegyu Lee

Dr. Daegyu Lee is a researcher at the Electronics and Telecommunications Research Institute (ETRI), Daejeon, South Korea. His research encompasses the design and validation of surveillance systems for the Korean Urban Air Mobility (K-UAM) project, vision-based navigation strategies for off-nominal scenarios, and a digital twin–based simulation frameworks for Advanced Air Mobility. He earned his B.S. degree in Automotive Engineering from Kookmin University in 2018, and his M.S. and Ph.D. degrees in Future Vehicle and Electrical Engineering from KAIST in 2020 and 2024, respectively. His interests include Detect-and-Avoid system architecture, non-cooperative aircraft detection, and simulation-based assessment methods such as Hardware-in-the-Loop and Pilot-in-the-Loop Simulations.