Saminda Wattuhewa, senior lecturer and course leader in cyber security, University of Portsmouth
A construction site is one of the worst places to rely on GPS. Concrete floors stack overhead, steel frames scatter the signal, and basements block it outright.
Yet this is exactly the environment where autonomous robots are increasingly expected to work precisely and navigate around moving crews, doing it without a satellite fix to tell them where they are.
The technology filling that gap is Simultaneous Localisation and Mapping (SLAM). Rather than relying on a pre-loaded map or an external positioning signal, a SLAM-equipped robot builds its own map in real time using cameras or Light Detection and Ranging (LiDAR), often combined with inertial measurements for extra stability between readings.
An inertial sensor alone has no way to sense the surrounding environment, so it supports the mapping rather than performing it on its own. LiDAR is the sensor this piece focuses on, since it is what the deployment described below actually uses.
The two problems – “where am I” and “what does this space look like” – are solved together rather than one depending on the other already being answered.
That combination matters more on a construction site than in most other places robots currently operate. Warehouse robots typically work in spaces designed to stay largely fixed between shifts.
A construction site is designed to do the opposite. Materials arrive and get moved. Floor plans get revised mid-build. Scaffolding goes up and comes down. A map that is accurate on Monday can be wrong by Wednesday, so the map itself has to be a living thing, not a one-off survey.
What SLAM actually does, and does not, do
It is worth being precise about what that live map actually buys you, because it is easy to overstate. SLAM’s job is localisation and mapping: working out where the robot is and what the space currently looks like.
It is not by itself what makes a robotic arm place a brick to a tight tolerance; that precision comes from the arm’s own control system and close-range sensing, a separate piece of engineering. What SLAM changes is what that arm is aiming at.
An autonomous cart or arm working from a live, corrected map is aiming at the site as it actually is today, rather than a static drawing that may already have been overtaken by a moved pallet or a revised wall line.
Get that target wrong and no amount of precision in the placement mechanism helps. It is worth being precise here too: SLAM on its own updates geometry rather than meaning; a purely geometric SLAM system has no concept of a beam being in the “wrong” place, only that the space now looks different from the last scan.
Semantic SLAM, a related research strand that labels mapped features with meaning rather than raw position alone, narrows this gap and is an active area of research, including a 2026 structured review specifically covering SLAM generation in the construction sector.
Recognising a change as a meaningful error, not just a difference, still generally needs the map compared against a design intent, typically a Building Information Model, plus planning logic to decide what to do about the mismatch. That comparison and decision layer sits on top of SLAM, not inside it.
Not a lab demonstration
Construction-site SLAM is an active area of applied robotics research in its own right, precisely because the standard assumptions behind SLAM, a mostly static scene and structure that does not change while you are mapping it, plus, for camera-based (visual) SLAM specifically, consistent lighting, break down on a real site.
A 2023 Purdue University thesis, for instance, focused specifically on adapting visual SLAM to handle the dynamic features and shifting structure of construction sites, rather than treating construction as just another indoor space to be mapped once.
It is also being tested outside the lab. Oxford Robotics Institute’s AutoInspect programme, which uses graph-based LiDAR SLAM to let mobile and legged robots map an environment and then inspect it autonomously, was deployed at Costain’s Gatwick Station construction site in autumn 2022, where it mapped the site and carried out autonomous inspection within four hours.
A separate, and importantly different, project at Virginia Tech put Boston Dynamics’ Spot robot to work on construction sites too, but as a remotely piloted platform for progress monitoring: an operator drives it remotely while it streams live 360-degree video, with an augmented-reality layer comparing what the camera sees against the planned Building Information Model.
It is not a SLAM deployment, but it is a useful contrast: it shows the design-model comparison described above already working in practice, just with a person doing the navigating rather than the robot doing it itself.
Why proponents think this matters
The wider case for adopting this now runs along three lines, and beyond the single observed fact Gatwick actually establishes, a full autonomous map-and-inspect cycle in four hours, the rest is worth treating as an argument rather than settled fact, since industry-wide data at scale is not yet public.
Safety is the most intuitive: autonomous survey and inspection work of the kind AutoInspect carried out at Gatwick can reduce the amount of time people spend in partially built, structurally uncertain spaces, particularly below grade, where GPS is denied outright, and at height, which carries its own fall and structural risk independent of GPS.
Productivity is the second, less obvious claim: the suggested gain is not that robots outpace people at any single task, but that they could work across shifts without the setup and re-survey overhead a changing site otherwise demands of a human crew re-establishing reference points by hand.
Gatwick’s four-hour result shows that cycle is achievable once; it does not by itself show sustained across-shift gains, which would need data from repeated use over time that has not been published.
The third is error reduction: material placed against a live, corrected map should in principle produce less rework than material placed against a drawing that has already been overtaken by events on site; this one remains a plausible extension of what SLAM does rather than something either named deployment above actually measured, and a careful reader should treat it that way until site-level data says otherwise.
What still gets in the way
None of that means the technology is a drop-in replacement for how construction currently runs, and an honest account of it should say so. LiDAR has two separate problems on a construction site, not one.
Airborne dust scatters and absorbs the laser pulses a scanner depends on, a well-documented failure mode in field robotics generally, degrading the return signal before it ever reaches a surface worth mapping.
Physical debris is a different issue: solid material blocking a clear line of sight, which occludes what is behind it rather than corrupting the signal itself. Both are common on an active site, and either can produce an incomplete or unreliable map.
Deployment cost and integration with existing site management software are also likely to remain real barriers for contractors who are not already running a robotics programme.
And a map that updates in real time is only as useful as the decisions built on top of it, which means the software layer interpreting that map, not just the sensor collecting it, is where a lot of the remaining engineering work sits.
Where this goes next
What is worth watching over the next few years is not whether a SLAM-equipped robot can work on a real construction site at all; Gatwick already answers that. The more useful question is which parts of the workflow adopt the technology first.
Oxford’s autonomous inspection and Virginia Tech’s teleoperated progress monitoring are related workflows, not two versions of the same task, and both are worth tracking on their own terms rather than assuming one is simply an early substitute for the other.
Site survey and progress monitoring are the easiest entry point. A robot’s live navigational map is not automatically survey-grade; formal surveys carry their own accuracy and completeness requirements that a navigation map is not built to meet.
But the underlying mapping process already captures much of the same positional and geometric data a survey needs, which shortens the path to that use case even if extra work remains to close the gap.
Material placement and repetitive assembly tasks are the harder, higher-value target, and the one where the gap between a research demonstration and a dependable piece of site equipment will actually get closed.
For an industry that has changed its core methods slowly for decades, a robot that can genuinely read a site as it changes, rather than one that needs the site to hold still for it, is a meaningfully different kind of tool.
About the author: Saminda Wattuhewa is senior lecturer and course leader in cyber security at the University of Portsmouth, with a research and teaching interest in robotics.
