Draft:NAVER LABS

• Comment: We are not interested in AI generated advert drivel Sulfurboy (talk) 06:47, 10 August 2026 (UTC)

---

NAVER LABS

Native name	네이버랩스
Type	Subsidiary
Industry	Research and development
Founded	January 2017
Headquarters	Seongnam , South Korea
Key people	Sangok Seok (CEO)
Parent	Naver Corporation
Website	www.naverlabs.com/en/

NAVER LABS (Korean: 네이버랩스; stylized as NAVER LABS) is a South Korean research and development company and a subsidiary of NAVER Corporation. It conducts research in robotics, spatial AI, physical AI, digital twins, and augmented and virtual reality technologies. It also operates NAVER LABS Europe, an artificial-intelligence research center in Grenoble, France.

NAVER LABS was established as a separate corporate entity in 2017 after operating as an internal research organization within NAVER. Sangok SEOK (Korean: 석상옥) serves as its chief executive officer.

History

Internal research organization (2013–2016)

NAVER established NAVER LABS as an internal research organization in 2013. It initially conducted research in artificial intelligence, autonomous driving, robotics, mapping, and human–machine interaction.

In 2016, the organization introduced M1, an indoor mapping robot designed to collect three-dimensional spatial data. It subsequently presented robotics platforms including the AROUND indoor autonomous robot and the AMBIDEX dual-arm robot. AROUND was later tested in a bookstore alongside AIRCART, an electrically assisted shopping cart.

Incorporation and European expansion (2017–2018)

NAVER LABS was incorporated as a separate NAVER subsidiary in January 2017.

In June 2017, NAVER acquired Xerox Research Centre Europe in Grenoble, France. The center was incorporated into NAVER’s research network and renamed NAVER LABS Europe, expanding the company’s research in computer vision, machine learning, natural-language processing, and related fields.

Development of cloud robotics and ARC (2019–2021)

At CES 2019, NAVER LABS and Qualcomm demonstrated a “brainless robot” system in which parts of the robot-control computation were transferred to cloud infrastructure over a 5G network. The project became a technical predecessor of the company’s cloud-robotics architecture.

In November 2020, NAVER LABS introduced ARC, an acronym for “AI, Robot, Cloud.” ARC was designed to coordinate localization, navigation, task execution, and infrastructure interaction for multiple robots through a shared cloud system.

Deployment at NAVER facilities (2022–2023)

NAVER’s 1784 headquarters began operating as the primary testbed for ARC-based robot services in 2022. Approximately 100 Rookie service robots were deployed to deliver parcels, meals, and beverages between floors.

NAVER also commercialized ARC eye, its cloud-based visual-localization system, through NAVER Cloud Platform.

In 2023, NAVER opened the GAK Sejong data center. GaRo, SeRo, and ALT-B were deployed to transport servers, manage IT assets, and carry workers around the site.

The same year, NAVER secured a five-year project to construct cloud-based digital-twin platforms for Riyadh, Makkah, Madinah, Jeddah, and Dammam in Saudi Arabia.

Platform and international expansion (2024–present)

In March 2024, NAVER LABS announced ARC mind, a web-platform-based robot operating-system architecture intended to simplify the development and deployment of robot services. ARC mind has been presented as part of the company’s longer-term platform strategy, although detailed information about public APIs and large-scale external deployments remains limited.

NAVER formally commenced its Saudi Arabian digital-twin project in July 2024. The project combined NAVER LABS’ mapping and spatial-intelligence technologies with NAVER Cloud’s infrastructure.

In June 2025, Team NAVER announced the completion of the initial digital-twin platforms for Makkah, Madinah, and Jeddah. The platforms cover more than 6,800 square kilometres and approximately 920,000 buildings.

In June 2026, NAVER LABS Europe released DIVINE, a universal AI encoder designed to reduce duplicated computation across multiple robot-perception tasks.

The following month, a robot-delivery service using NAVER’s digital-twin and robotics technologies began full-scale operation at Tokyo Midtown Yaesu in Japan.

Main research areas

Physical AI

NAVER LABS uses the term physical AI for systems that perceive physical spaces, make decisions from sensor and spatial data, and translate those decisions into robot or machine actions. Its approach combines real-world operational data, cloud robotics, and onboard AI.

Data collected from offices, data centers, and other testbeds is used to improve robot perception and behavior. ARC connects robots, spatial information, and building infrastructure through cloud services, while lightweight onboard models handle tasks that require low latency or continued operation when network conditions are unstable.

ARC

ARC is NAVER LABS’ cloud-based multi-robot intelligence system. The name is derived from “AI, Robot, Cloud.” It coordinates multiple robots and connects them to digital-twin data, service applications, elevators, security doors, and other building infrastructure.

ARC’s functions have been presented under the labels ARC brain, ARC eye, and ARC mind. These labels describe related functions rather than completely independent products:

• ARC brain covers robot-fleet management, task allocation, route planning, and coordination with spatial and service infrastructure.

• ARC eye combines digital-twin data with visual-localization technology to estimate the position and orientation of robots or camera-equipped devices.

• ARC mind is a web-platform-based robot OS architecture intended to make robot-service development and deployment more accessible to web developers.

The system was first deployed on a large scale at NAVER 1784 and was later adapted to the different robot types used at GAK Sejong. NAVER has since described ARC as a physical execution layer through which AI-agent plans could be converted into tasks performed by robots, although this remains a developing use case.

Robotics

NAVER LABS conducts research in autonomous navigation, mobile manipulation, robot learning, force control, human–robot interaction, and multi-robot coordination. It develops both service robots intended for human environments and specialized robots for industrial facilities.

The company uses its robots as research and operational platforms rather than treating each robot as an isolated product. Navigation, task planning, spatial data, service interfaces, and software updates can be distributed between onboard computers and ARC.

Major robots

AROUND series

AROUND is a wheeled indoor autonomous-robot platform introduced for operation in public and commercial spaces. The platform was tested in bookstores and other environments containing pedestrians and changing obstacles. Later variants included AROUND C, which was used for research into socially acceptable navigation and delivery behavior.

Rookie and Rookie 2

Rookie is an autonomous service robot deployed at NAVER 1784. Connected to ARC, it moves between floors using robot-only elevators and performs parcel, meal, and beverage deliveries. Approximately 100 Rookie robots have operated in the building.

Rookie 2 is a later platform developed to support expanded smart-building services and physical-AI research. It retains the cloud-connected operating structure while placing greater emphasis on local perception and AI functions.

GaRo, SeRo, and ALT-B

GaRo, SeRo, and ALT-B were developed for GAK Sejong.

• GaRo is an autonomous transport robot that carries servers and other heavy IT assets between storage and server facilities. It can transport loads of up to 400 kilograms.

• SeRo is an automated asset-management robot that stores, retrieves, and records server assets in the data center’s IT warehouse.

• ALT-B is a driverless autonomous shuttle that carries workers between major locations on the data-center campus.

The three systems are connected to ARC, allowing different robot types and service functions to be coordinated through a shared platform.

NURI

NURI is an outdoor mobile robot developed for autonomous operation beyond indoor smart-building environments. A mission demonstration involving NURI was publicly presented at NAVER 1784 in May 2026 during a visit by South Korea’s Minister of Land, Infrastructure and Transport. Publicly available technical specifications and deployment plans remain limited.

AMBIDEX

AMBIDEX is a dual-arm robot designed for safe interaction with people and for research into manipulation. Unlike conventional industrial arms with motors located at each joint, AMBIDEX uses a cable-driven mechanism that reduces the weight of its moving arms while allowing rapid movement, force control, and the handling of relatively heavy objects.

The system has been used in teleoperation, force-control, imitation-learning, and task-learning research. Demonstrations have included cooking, drawing, and other manipulation tasks based on motions collected from human operators.

Mini Cheetah

NAVER LABS uses the MIT Mini Cheetah quadruped platform for research into dynamic locomotion, reinforcement learning, and physical-AI control. Mini Cheetah was originally developed at the Massachusetts Institute of Technology and is therefore a research platform used by NAVER LABS rather than a proprietary NAVER robot design.

Spatial AI

NAVER LABS defines spatial AI as technologies that allow machines to recognize and reason about three-dimensional environments, including the geometry of spaces and the positions and relationships of people, objects, and cameras. Digital twins form part of this spatial-AI framework by representing physical environments as structured three-dimensional data.

Its research covers mapping, visual localization, 3D reconstruction, novel-view synthesis, human modeling, and foundation models for spatial perception. These technologies enable robots and other next-generation platforms to understand physical environments and provide services within them.

Mapping devices

R series

R1 is a vehicle-based mobile mapping system developed for road-data collection and hybrid HD mapping. It combines cameras, two- and three-dimensional LiDAR, GPS, an inertial measurement unit, a fibre-optic gyroscope, and wheel encoders. Data collected by R1 is combined with road information extracted from aerial imagery to construct high-definition maps.

M series

The M series consists of 3D mapping robots designed to efficiently construct digital-twin data for large-scale indoor spaces. M1 was introduced in 2016, followed by the upgraded M1X and M2. These robots collect point-cloud and image data that can be processed into maps for robot navigation, visual localization, digital twins, and indoor augmented-reality services.

M2 extended the platform for more varied indoor and outdoor environments and incorporated a vertical lifting mechanism to collect spatial data from different heights.

T series

The T series consists of portable mapping devices intended for spaces that are difficult to cover with large mapping robots or vehicles. The devices allow an operator to collect images and three-dimensional spatial data while moving through indoor and outdoor environments. T3 was designed to reduce equipment weight while increasing the range of spaces that could be mapped.

P series

P1 is a panoramic mobile mapping system that captures precise urban spatial data through NAVER Street View vehicles. Its data has been used to create panoramic three-dimensional content, including NAVER Map’s Street View 3D service.

Core technologies

Visual localization

Visual localization determines the position and orientation of a camera by comparing an input image with a previously constructed spatial map. It can provide positioning in indoor environments where satellite navigation is unavailable.

NAVER LABS combines visual localization with visual–inertial odometry and object tracking. The technology has been applied to robot positioning at 1784 and to indoor augmented-reality navigation services, including a demonstration at COEX in Seoul.

Novel view synthesis

Novel view synthesis, or NVS, creates images of a scene from viewpoints that were not directly captured. NAVER LABS researches NVS for realistic map visualization, urban digital twins, spatial content, and virtual camera movement.

The technology has been used in NAVER Map’s Flying View 3D to generate smoother and more realistic views around city landmarks.

DUSt3R

DUSt3R, short for Dense and Unconstrained Stereo 3D Reconstruction, is a 3D vision model developed by NAVER LABS Europe. It estimates three-dimensional scene structure directly from images by predicting a pointmap in which each image pixel is associated with a three-dimensional coordinate.

Unlike conventional structure-from-motion pipelines, DUSt3R does not require camera calibration parameters or camera poses as initial inputs. Its principal configuration takes two images and can infer their three-dimensional relationship even when their visible regions do not overlap. Its learned representation can also support monocular depth estimation, although a single-image depth result is not equivalent to a complete multi-view reconstruction.

DUSt3R provides a common representation for depth estimation, camera-pose estimation, pixel correspondence, and 3D reconstruction. It replaces several separately engineered stages of conventional pipelines, but does not eliminate uncertainty or reconstruction error.

DUSt3R belongs to a broader research line that includes CroCo, MASt3R, MASt3R-SfM, Pow3R, MUSt3R, and PanSt3R. These are separate models with different architectures and research objectives rather than versions released under a single DUSt3R product name.

DIVINE

Developed and released by NAVER LABS Europe in June 2026, DIVINE is a universal AI encoder for robot perception. Conventional multi-model robot systems may use a separate encoder for each visual task, causing the same camera input to be processed repeatedly. DIVINE instead produces a shared representation that can be passed to task-specific decoders.

The encoder is trained through multi-teacher knowledge distillation. DINOv2 supplies two-dimensional visual representations, MASt3R—an extension of DUSt3R—supplies three-dimensional spatial representations, and Multi-HMR supplies representations related to three-dimensional human understanding. During inference, the input is encoded once and shared by the different decoders.

NAVER reported that, in its evaluated configuration, DIVINE reduced encoder memory use by approximately 90%, increased encoding speed by up to 12 times, reduced total system memory use by approximately 62%, and improved complete-system processing speed by up to four times. These figures describe the reported experimental system and should not be treated as universal performance guarantees.

Its currently demonstrated applications are primarily computer-vision tasks. NAVER LABS Europe has described future research aimed at extending the architecture to multimodal tasks such as visual question answering.

Applications

NAVER LABS’ spatial-intelligence technologies have been applied to NAVER Map Street View 3D, Store View 3D, Flying View 3D, indoor augmented-reality navigation, and virtual-tour content for NAVER Pay Real Estate. The examples below describe how mapping, localization, digital-twin, and novel-view-synthesis technologies are used in services.

NAVER Map Street View 3D

Street View 3D combines panoramic imagery with spatial data collected using P1. It adds three-dimensional geometry to street-level imagery so that map labels, advertisements, and other digital content can be positioned relative to the surrounding scene.

Indoor augmented-reality navigation

NAVER LABS’ visual-localization technology has been applied to indoor AR navigation. A smartphone camera image is compared with a preconstructed feature map to estimate the user’s location and viewing direction, allowing navigation information to be aligned with an indoor environment.

Flying View 3D

Flying View 3D presents city landmarks from continuously changing virtual viewpoints. The service combines digital-twin data and novel-view-synthesis techniques to generate perspectives that are difficult to obtain through ordinary street imagery.

Other applications

NAVER LABS’ spatial data has also been used for NAVER Real Estate virtual tours, robot navigation, smart-building operation, high-definition road maps, urban-planning simulations, and visual-effects production. Urban models produced with ALIKE were used as reference material in visual-effects work for the second season of the Netflix series Sweet Home.

Digital twins

Within its spatial-AI research, NAVER LABS develops digital-twin technologies that reconstruct buildings, roads, and cities as three-dimensional datasets. The resulting models can be connected with administrative, environmental, traffic, and infrastructure data for visualization and simulation.

Its digital-twin work combines aerial photography, mobile mapping, automated 3D modeling, visual localization, and cloud infrastructure. Applications include city planning, flood-risk analysis, road and building management, robot navigation, augmented reality, and realistic map content.

ALIKE

ALIKE is NAVER LABS’ city-scale digital-twin solution. It uses aerial imagery, artificial intelligence, large-scale distributed processing, and mapping data to generate three-dimensional city models, road layouts, and high-definition maps.

ALIKE automates parts of the modeling process that would otherwise require extensive manual processing. Its components can be used separately or together depending on whether a project requires visual 3D models, road geometry, or autonomous-driving maps.

Unlike workflows centered on manually modeling individual buildings, ALIKE is designed to automate large portions of city-scale reconstruction. It can produce visual 3D models, road layouts, and high-definition maps from a shared set of aerial images and mapping data, allowing its components to be selected according to the requirements of a project.

Major achievements

NAVER 1784

NAVER 1784 is NAVER’s second headquarters in Seongnam and NAVER LABS’ principal smart-building testbed. The building integrates robotics, AI, digital twins, cloud computing, private 5G networking, and building infrastructure.

Approximately 100 Rookie robots provide delivery services across multiple floors. ARC coordinates robot positions, routes, and tasks and connects the robots to security doors and robot-only elevators. The building also provides operating data for robot navigation, human–robot interaction, fleet management, and smart-building services. NAVER describes 1784 as the world’s first robot-friendly high-rise office building.

GAK Sejong data center

GAK Sejong is NAVER’s hyperscale data center in Sejong City. Whereas 1784 focuses on service robots operating around office workers, GAK Sejong is used to test robots and automation systems for large industrial facilities.

GaRo transports heavy servers and IT assets, SeRo manages storage and retrieval, and ALT-B carries workers between locations on the campus. ARC coordinates the different robot types and connects their operations with the data center’s asset-management systems.

Tokyo Midtown Yaesu

Tokyo Midtown Yaesu became NAVER’s first major overseas reference for the smart-building model developed at 1784. The project was carried out with Mitsui Fudosan, NTT East, and NAVER Cloud.

After pilot operation, a full-scale delivery service was announced in July 2026. Four delivery robots transport food and beverages from restaurants on the basement first and fifth floors to an office area on the seventh floor. The system uses a cloud-based digital twin and coordinates the robots with elevators and security doors. Users place orders through LINE.

Makkah, Madinah, and Jeddah

Under a five-year project signed with Saudi Arabia’s Ministry of Municipal, Rural Affairs and Housing in 2023, Team NAVER began constructing and operating cloud-based digital-twin platforms for Riyadh, Makkah, Madinah, Jeddah, and Dammam. The project entered its construction phase in July 2024, with NAVER LABS supplying spatial-intelligence and digital-twin technologies and NAVER Cloud providing cloud infrastructure.

The first phase, covering Makkah, Madinah, and Jeddah, was completed in June 2025. At the completion ceremony, NAVER unveiled the KSA Digital Twin System, a Saudi Arabian digital-twin platform based on actual city data. The models cover more than 6,800 square kilometres and approximately 920,000 buildings, while the platform supports terrain and skyline analysis, building-code review, urban planning, flood-risk visualization, and weather-data integration.

New Murabba

In 2025, NAVER Cloud and New Murabba Development Company announced a technology partnership related to the New Murabba urban-development project in Riyadh. The parties identified possible areas of cooperation including digital twins, cloud infrastructure, robotics, autonomous systems, and smart-city services.

The agreement should be described as a partnership or memorandum of understanding rather than as a completed NAVER LABS deployment. Specific systems, operating scope, and implementation schedules had not been publicly confirmed at the time of the announcement.

NAVER LABS Europe

NAVER LABS Europe is NAVER’s European artificial intelligence research center in Grenoble, France. Its predecessor, Xerox Research Centre Europe, was established in 1993. NAVER acquired the center in 2017 and incorporated it into the NAVER LABS research network.

The center organizes its robotics-oriented research around three broad capabilities: perception of environments and people, interaction with humans and other systems, and action or sequential decision-making. Its research fields include computer vision, natural-language processing, robot learning, human–robot interaction, reinforcement learning, machine learning for optimization, information retrieval, and systemic AI.

Its spatial-AI research includes the CroCo, DUSt3R, MASt3R, MASt3R-SfM, Pow3R, MUSt3R, and PanSt3R model families. Research on universal representations includes UNIC, DUNE, and DIVINE. DUNE uses DINOv2, MASt3R, and Multi-HMR as teacher models to combine two-dimensional vision, three-dimensional spatial understanding, and human understanding in a common encoder.

The center also operates the HUMANS research program, which develops human- and spatial-understanding technologies for robot perception and interaction. Its research includes three-dimensional human pose and shape estimation, reconstruction of people within 3D scenes, and representations of varied human body types.

Anny is a parametric 3D human-body model designed to represent body shapes across different ages, heights, and weights, from infants to older adults. Its body-shape distributions incorporate World Health Organization population data. The Anny model and code were publicly released in November 2025 under the Apache 2.0 license.

Anny-One is a synthetic human dataset generated using Anny. Released in February 2026, it contains 780,964 photorealistic multi-view and multi-person images generated from 22,951 synthetic indoor scenes. It was developed to supplement real human datasets, which are limited by privacy, annotation costs, demographic coverage, and diversity. The dataset is available for non-commercial use under the CC BY-NC-SA 4.0 license.

NAVER LABS Europe publishes research through international conferences and releases selected papers, code, models, and datasets. Its work is used both as basic research and as a source of perception, planning, and interaction technologies for NAVER LABS’ robotics and spatial-AI systems.

See also

• Naver Corporation
• Sangok Seok
• Cloud robotics
• Digital twin
• Computer vision
• Augmented reality