South Korean construction conglomerate Lotte Construction announced on July 29 that it has moved an autonomous floor-cleaning robot from proof-of-concept testing to full operational use at an active apartment construction site in Incheon — a harder deployment environment than any comparable Korean robot cleaning program to date, and one that runs without a human operator even after the lights go out.
The robot, developed with Bucheon-based robotics startup TXR Robotics, navigates by consulting a pre-loaded map of its work area rather than building a spatial model from scratch in real time. That design choice — trading real-time mapping flexibility for reliable, repeatable localization — directly addresses the most documented failure mode of autonomous robots on construction sites, and it is why the machine can sweep dust and collect debris through the night shift without a human minder.
(Exchange rate as of July 29, 2026; conversions are approximate.)
Korean Construction’s Dustiest, Most Dangerous Chore Gets a Robot
Construction site cleaning has historically been one of the physically harshest tasks in the industry. Workers have manually collected debris and removed fine dust with vacuum equipment — a process that generates large quantities of airborne particulate, forcing workers to wear dust-proof respirator masks throughout, even in summer heat. Fine particulate exposure is a chronic occupational health hazard in confined construction environments.
Lotte Construction cited the relief of this specific burden as the primary rationale for the deployment. An official company statement said that automating repetitive cleaning tasks is expected to allow site workers to focus more fully on core construction work and improve overall site efficiency.
The company completed a proof-of-concept trial with TXR Robotics at the Incheon Hyosung District Urban Development apartment project — specifically the 4BL block currently under active construction — before transitioning the robot to operational duty.
How the Robot Finds Its Way in the Dark: Pre-Mapped Navigation vs. Real-Time SLAM
The technical architecture at the center of this deployment is not what most readers would assume from a “smart robot” description. Many autonomous robot systems rely on Simultaneous Localization and Mapping — SLAM — an approach in which the machine builds its map of the environment at the same time as it navigates through it. SLAM is mathematically elegant and enables robots to function in entirely unknown spaces, but it is notoriously fragile in construction environments for a specific and documented set of reasons: building sites combine low ambient illumination, low-texture surfaces such as bare concrete, and low structural differentiation — floors, ceilings, and walls that look nearly identical to a camera sensor — creating conditions under which the feature-matching algorithms that SLAM depends on routinely fail.
Recent academic research in a peer-reviewed construction engineering journal found that SLAM in large-scale public building construction sites is “a highly challenging task” due to “complex spatial structures, time-varying conditions, and dynamic uncertainties” according to a Springer Nature study. A separate study found that visual SLAM localization errors in uneven low-light construction environments can be reduced by more than 65% with specialized deep-learning enhancement algorithms — but only after applying those algorithms; without them, performance degrades significantly.
TXR Robotics’ cleaning robot sidesteps these problems entirely by loading a pre-built map of the cleaning zone before operations begin. Rather than attempting to construct a spatial model in real time under shifting, low-light conditions, the robot learns the spatial layout and cleaning coverage of its designated area from a pre-loaded site map, then navigates that zone autonomously using the map as its reference frame. This is the same architectural approach that won the DARPA Urban Challenge for Carnegie Mellon’s Boss and Stanford’s Junior in autonomous vehicle research — both pre-built-map systems outperformed real-time SLAM approaches in competitive conditions.
The robot’s obstacle-detection system operates as a separate layer on top of this global navigation framework: onboard sensors detect workers and physical obstacles in the machine’s path and route around them dynamically, without requiring a map update. Onboard lighting enables the same localization to function in the dark.
The practical tradeoff is worth naming. A pre-loaded map approach is less adaptable than full SLAM: if the construction site layout changes significantly — materials stacked in new locations, walls closed off — the robot’s operational zone must be re-mapped. For a construction site, where the layout evolves continuously, this is a real constraint. Whether Lotte and TXR have a workflow to handle map updates as construction progresses is not disclosed.
What Sets This Deployment Apart From Korea’s Prior Construction Robot Programs
POSCO E&C, another major Korean builder, announced a comparable autonomous cleaning robot deployment in February 2026 in partnership with robotics company Clobot. That deployment is at the Shinbanpo 21st housing reconstruction project — but specifically in the post-construction phase, cleaning community spaces in a completed residential complex. Elevators, corridors, and communal halls have stable, unchanging layouts: the navigation challenges are substantially lower than in an active construction zone.
Lotte Construction’s Incheon deployment targets the active construction phase itself — the period when floors are incomplete, materials are stacked unpredictably, and human traffic is dense and variable. This is a harder operational problem, and it is why the engineering choice to use a pre-loaded map matters: real-time SLAM is precisely the approach that struggles most in exactly this kind of environment.
TXR Robotics, the startup that built the cleaning machine, listed on the Korean stock exchange in March 2025 after founding in 2024. The company’s product page specifies dual cleaning modes for fine dust and debris, patrol cleaning mode for autonomous route-following, onboard obstacle detection and avoidance, and the ability to remotely monitor the machine and upload real-time cleaning data.
South Korea’s broader push to automate physically demanding and repetitive construction work is driven in part by a structural demographic reality. Korea has the world’s lowest national fertility rate — 0.7 live births per woman as of 2023 — and a rapidly aging construction workforce. The country’s government has explicitly promoted a “K-robot economy” strategy targeting deployment of robots across manufacturing, construction, and services.
The global commercial cleaning robot market — of which construction-phase cleaning is a specialized subset — is projected to expand from approximately $14.8 billion in 2025 to nearly $96.8 billion by 2035, a compound annual growth rate of 20.6%, driven largely by rising labor costs and demand for consistent autonomous performance in industrial settings.
Lotte’s Wider Startup Automation Program
The cleaning robot deployment is one element of a broader innovation strategy. In July 2026, Lotte Construction separately secured up to 620 million Korean won (approximately $420,000 USD) in proof-of-concept support funding through open innovation programs run by the Korea Institute of Startup and Entrepreneurship Development and the Seoul Business Agency. That funding supports five collaborative PoC projects with different startups, including construction automation startup Spanner, an AI-based supervision records management company called Double S&DT, defect-detection provider Frictech, and MobilityOne, which focuses on integrated management of heterogeneous robot fleets.
Lotte Construction established a dedicated startup collaboration unit within its Technology Research Institute in 2023 and has indicated that technologies proven in PoC programs will be candidates for joint development and joint IP applications.
The company also plans to expand the autonomous cleaning robot’s use into final-stage pre-handover cleaning operations — the meticulous cleanup required when a building is complete but before residents move in. That phase involves more controlled and stable layouts than active construction, which would make it an easier operating environment for the current pre-loaded-map navigation system.
What Comes After Pre-Loaded Maps
The real test for construction robotics is not a controlled floor-cleaning deployment — it is whether autonomous systems can adapt to the continuous layout changes that define active building sites. Pre-loaded-map navigation solves today’s deployment problem practically and reliably, but it defers the harder question: what happens when the map goes stale halfway through a floor’s construction?
The research community is actively working on that problem. A recent Springer Nature paper proposed a systematic evaluation framework for LiDAR SLAM algorithms in large-scale public building construction sites — specifically because no standardized dataset for that environment previously existed, which had “considerably slowed the progress of SLAM research for construction robots.” LiDAR SLAM, which uses laser rangefinding rather than cameras, is significantly more robust to the illumination changes that defeat visual SLAM — and it is likely the direction in which construction-site navigation systems will evolve as costs fall.
For now, Lotte Construction and TXR Robotics have demonstrated that a practically scoped autonomous cleaning robot can operate reliably on an active Korean construction site, reduce dust exposure for workers, and run through the night without human attendance. Whether SLAM-based systems can eventually take over the work of robots like this — adapting to layouts as they change rather than relying on a pre-built map — is the engineering question that will determine the ceiling of what autonomous construction-site robots can ultimately do.
Frequently Asked QuestionsWhat makes deploying an autonomous cleaning robot on an active construction site harder than in a warehouse or post-construction building?
Active construction sites combine three conditions that defeat most standard autonomous navigation systems: low ambient lighting, low-texture surfaces (bare concrete walls and floors that look nearly identical to camera sensors), and dynamic, constantly changing layouts as construction progresses. Warehouses and completed buildings have stable layouts, consistent lighting, and clear visual landmarks. The pre-loaded-map navigation approach used in the Lotte Construction deployment addresses the illumination and texture problems but cannot fully solve the layout-change problem — if the site reconfigures significantly, the map must be updated.
Why does the robot use a pre-loaded map instead of real-time SLAM, and what does that tradeoff mean in practice?
Simultaneous Localization and Mapping (SLAM) — where a robot builds its map while navigating — is the more flexible approach, but it is prone to failure in the low-light, low-texture conditions of construction sites, where feature-matching algorithms often cannot find consistent visual anchors. A pre-loaded map gives the robot a reliable reference frame without depending on real-time feature detection. The tradeoff is adaptability: if the site layout changes substantially, the robot’s operational zone needs to be re-mapped. For a bounded, repeating cleaning task in a known zone, the pre-loaded-map approach is more reliable. For a robot that needs to explore unfamiliar spaces or adapt to daily layout changes, real-time SLAM remains the target architecture.
How does the robot avoid running into construction workers?
Obstacle detection and avoidance operates as a separate layer from the robot’s global navigation. Onboard sensors — the specific sensor type (LiDAR, ultrasound, depth camera, or a combination) is not disclosed in TXR Robotics’ public specifications — continuously scan the robot’s path for workers, equipment, and debris piles. When an obstacle is detected, the robot routes around it without updating its underlying pre-loaded map. This local avoidance system allows the robot to function safely in an environment with unpredictable human traffic while still following its pre-planned cleaning route.
How does this deployment fit into South Korea’s broader construction automation picture?
Korea’s construction sector is automating under a combination of structural labor shortage — the country has the world’s lowest national fertility rate — and explicit government support for robotics across industry sectors. POSCO E&C, Hyundai, and now Lotte Construction have all announced robotic deployments on or around construction sites in 2026. The Lotte/TXR deployment is notable for targeting the active construction phase rather than post-construction maintenance, which is the harder technical context. Government programs through KISED and the Seoul Business Agency are providing proof-of-concept funding for exactly this kind of startup-chaebol technology validation partnership.