Industries / Construction and property / Commercial and residential construction / ConTech
Recording site progress
Recording site progress means capturing what has actually been built, with cameras, laser scanners, robots and software, and comparing it with the plan, the schedule and the building model.
7
Milestones achieved, from 1998 to 2021.
3
To watch for: announced, not yet confirmed outside.
3
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1998
A scanner records as-built
In June 1998 Cyra Technologies publicly launched the Cyrax 2400 at the AEC Systems trade show in Chicago. It is a laser scanner: it sweeps a laser beam across a scene, measures the distance to each spot the beam hits, and saves the results as a point cloud, a dense set of 3D points that records the shape of what is there. Cyra aimed it at "as-built and site surveys", and it weighed more than 100 pounds. Leica Geosystems completed its purchase of Cyra in February 2001. In 2006 a US General Services Administration (GSA) official called laser scanning "one of the more mature tools" in the agency's building-model program, used to give designers reliable records of existing buildings.
Credit
Ben Kacyra and Cyra Technologies; Leica Geosystems
2009
Site photos meet the plan
In June 2009 Mani Golparvar-Fard and Feniosky Peña-Mora of the University of Illinois and Silvio Savarese of the University of Michigan published D4AR. It took a site's unsorted daily progress photos, worked out where each was taken from, and rebuilt a rough 3D model of the site. It then laid that model over the 4D model, the 3D design linked to the schedule, so that planned and built work could be compared, and showed the results in traffic-light colors. The authors reported early results from three projects. Golparvar-Fard, Derek Hoiem and Tim Bretl, all Illinois professors, founded the company Reconstruct. By the university's account it had more than 1,000 projects underway in 2022, including work with Japan's Taisei Corporation on the North-South Commuter Railway in greater Manila.
Credit
Golparvar-Fard, Peña-Mora and Savarese; Reconstruct
2015
Cameras miss hidden work
A Technical University of Munich team reported that "due to the numerous obstructions found on a construction site", "only a minority of building elements can be detected directly" in 3D models built from site photos. They used the order of work in the schedule to infer elements the cameras could not see, since, for example, a slab cannot be poured before the columns that carry it. Ten years later, OpenSpace described its progress-tracking product as combining jobsite imagery "with computer vision and expert human verification".
LessonA camera confirms only the work it can see, so progress tools have to reason about hidden work or send a person to check.
Credit
Alexander Braun, Sebastian Tuttas, André Borrmann and Uwe Stilla (ISARC 2015, Oulu, Finland)
2018
Hard-hat cameras map the site
In June 2018 OpenSpace launched a product in which a worker wears a consumer 360-degree camera on a hard hat and walks the site. Its software stitches the footage into a street-view-style record within a day, so any spot can be compared over time. MIT News reported in October 2020 that the software places the photos on the work plans automatically, and that OpenSpace had mapped more than 1.5 billion square feet of projects. By May 2025 the US military's shared guide specification for photo records (UFGS 01 32 33.00 10, used by the Army Corps of Engineers, NAVFAC and AFCEC) lets a project office require "360 photographs" of interior work before walls, floors and ceilings are closed, indexed on interactive drawings, with a view that overlays an image on the building model.
Credit
OpenSpace (Jeevan Kalanithi, Philip DeCamp, Michael Fleischman); USACE, NAVFAC and AFCEC
2019
A site diary falls short
In White Constructions v PBS Holdings, the Supreme Court of New South Wales dismissed White's claim for A$1,935,199 against PBS, the firm it had hired to design the sewers for a subdivision, for delay it blamed on the design's late approval. The site diary was "comprehensive" and well kept, and it repeatedly recorded "Waiting for approved sewer design". But the judge found that "the diary does not identify the activities, if any, which were being adversely affected by the wait." He also held that neither side's delay expert used a method suited to the case.
LessonA progress record proves a delay only if it shows which work was held up, where and when.
Credit
Justice Hammerschlag, Supreme Court of New South Wales, 6 September 2019
2020
A robot dog walks sites
In January 2020 the Canadian contractor Pomerleau began a six-month trial of Boston Dynamics' four-legged robot Spot on a site near Montreal. Spot carried a 360-degree camera linked to the HoloBuilder progress software. Pomerleau's Eric Lessard told ENR in May 2020 that Spot let the company "reallocate 20 hours per week of labor" by taking site photos after hours. He also said the spacing of columns and walls on the 500,000 sq ft site was too great for Spot to navigate by its onboard cameras alone, so his team added a lidar scanner, and that snow caused trouble. In the same article Trimble described Hensel Phelps collecting scans of its Denver International Airport renovation with a Trimble X7 scanner on Spot's back and comparing them with the building model.
LessonA robot can record only a site it can find its way around, and on a large site that took an extra scanner.
Credit
Pomerleau; Hensel Phelps; Boston Dynamics; Trimble; HoloBuilder
2021
Helmet cameras check the model
Buildots, founded in 2018 by Roy Danon, Aviv Leibovici and Yakir Sudry, checks 360-degree images from cameras on hard hats against the building model to track what has been installed. Build Group, its first US customer, used it on the 19-story M2 tower in San Francisco, with two engineers walking the site with the cameras. Build Group's vice president of operations, Chad Krause, told ENR in August 2021 that "it took three or four months to collaborate back and forth and make sure the models got set up correctly."
LessonAn automatic progress check is only as ready as the model behind it, and readying the model took months.
Credit
Buildots; Build Group
Watch for
Progress checked by people
OpenSpace bought the progress-tracking firm Disperse in October 2025 and sells the joint service, OpenSpace Progress Tracking, as an add-on that "combines jobsite imagery with computer vision and expert human verification" and links to scheduling software such as Oracle Primavera P6. In May 2026 OpenSpace wrote that its human verification layer "results in 95% accuracy", without saying how this was measured. Company-wide, OpenSpace reports more than 75,000 projects in 124 countries.
Achieved whenA contractor, owner or researcher compares its percent-complete figures with a hand count on the same project and publishes the error.
Claimed by
OpenSpace, October 2025 and May 2026
Delays cut by half
Buildots says its platform, which checks site photos against the schedule, "has been proven to reduce project delays by up to 50%", which it equates to two to three months on average projects. It cites no study, and gives no project count beyond "hundreds of projects globally".
Achieved whenAn outside party publishes delay figures for comparable projects with and without the platform, with their baselines.
Claimed by
Buildots, April 2026
Utility progress from drone photos
Buildots says a service launched in October 2025 turns processed drone photos of underground utility work into progress percentages and pace forecasts, and calls it "the first to transform drone-captured imagery into actionable intelligence". The contractor Mortenson piloted it on a data center project in the southern US. No accuracy or speed figures were given.
Achieved whenMortenson or another contractor publishes the service's progress figures checked against survey or as-built records.
Claimed by
Buildots, October 2025
Open
01
A percent-complete answer key
A public test set for progress-tracking tools. Walk one or more floors of a building under construction each week with a 360-degree camera. For a fixed list of elements taken from the building model, such as each stud wall, run of drywall, ceiling grid bay, duct run and sprinkler drop, have a person record whether it is installed, checked on site. Then run open methods, and any commercial tool whose user agrees, on the same walks. Publish how far each method's percent complete, the share of planned work that is installed, is from the hand count.
The briefCredit · unclaimed02
Phone check before a pour
A phone tool that checks embed plates and pipe sleeves in formwork against the building model before concrete is poured, and flags any item out of tolerance. Embed plates are steel plates cast into concrete so that steel or equipment can be fixed to it later; sleeves are short tubes that leave holes for pipes and cables. Build the test on mock forms on the ground, made of plywood, rebar offcuts, plates and plastic sleeves, with no concrete.
The briefCredit · unclaimed03
Activity dates from site cameras
A tool that reads the image archive of a fixed site camera and writes down the date each visible activity started and finished, such as each floor's slab, the steel frame, facade panels and the roof. The result is an as-built schedule, the record of what actually happened when. Test it against dates a person reads from the same images and against the contractor's daily reports.
The briefCredit · unclaimed
Open 01 · the brief
A percent-complete answer key
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A public test set for progress-tracking tools. Walk one or more floors of a building under construction each week with a 360-degree camera. For a fixed list of elements taken from the building model, such as each stud wall, run of drywall, ceiling grid bay, duct run and sprinkler drop, have a person record whether it is installed, checked on site. Then run open methods, and any commercial tool whose user agrees, on the same walks. Publish how far each method's percent complete, the share of planned work that is installed, is from the hand count.
Why it is still open
A December 2025 review of 51 open image and scan datasets for construction monitoring found 9 that support progress tracking. It called for datasets "meaningfully linked with BIM models, schedules, and digital twins", and said "no unified or widely adopted protocols exist for standardized evaluation." This research found no independent comparison of progress-tracking tools against a hand count. The accuracy figures vendors publish come from their own tests (see Watch for).
Built with what exists
A consumer 360-degree camera such as an Insta360 X4 or Ricoh Theta; the building's IFC model (the open file format for building models), read with the open IfcOpenShell library; COLMAP, open software that works out where each photo was taken; Meta's open SAM 2 to outline objects in the photos; a model that answers fixed yes-or-no questions about an image, such as Gemini Decision 1 (see the Note "A thousand questions, one pass"), or a small open vision model fine-tuned on one graphics chip, such as Liquid AI's LFM3 (see the Note "One GPU to fine-tune"). Workers will appear in the photos, so get the contractor's permission, tell the crews, and blur faces before anything is released.
First step
Ask one contractor or a university building project for permission to walk one floor weekly, and pull that floor's element list from its model with IfcOpenShell.
Done looks like
Eight weekly walks of at least 2,000 m² of interior work, with an answer key for at least 500 elements per walk released under an open license. A published table gives each method's percent-complete error by trade and by room, and its count of elements wrongly called installed or missing.
Open 02 · the brief
Phone check before a pour
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A phone tool that checks embed plates and pipe sleeves in formwork against the building model before concrete is poured, and flags any item out of tolerance. Embed plates are steel plates cast into concrete so that steel or equipment can be fixed to it later; sleeves are short tubes that leave holes for pipes and cables. Build the test on mock forms on the ground, made of plywood, rebar offcuts, plates and plastic sleeves, with no concrete.
Why it is still open
A 2020 paper from California Polytechnic State University says checking embed positions "is completed manually", cites a tolerance from ACI 117 of 1 in (25 mm) either way, and tested its own scan check only in a laboratory, with a FARO Focus 350 laser scanner. In a 2022 indoor test of an iPad Pro's lidar sensor, three quarters of the iPad's points lay within 6.6 cm of the reference model, against 2.8 cm for photogrammetry (measuring from many overlapping photos); both are wider than that tolerance. This research found no published test of whether a phone can catch a misplaced plate or sleeve.
Built with what exists
An iPhone Pro or iPad Pro with lidar and a scanning app such as Scaniverse, Polycam or SiteScape; COLMAP for photogrammetry from the phone's photos; printed AprilTag targets for scale and alignment; the open Open3D library to line the scan up with the model; IfcOpenShell to read each item's planned position from an IFC model; a steel tape and a laser distance meter for the answer key. Lidar in phones and tablets, with free scanning apps, is the advance that makes a cheap test possible.
First step
Build one 2.4 m by 1.2 m plywood mock-up with 20 plates and sleeves, tape each position, and scan it with two apps.
Done looks like
Three mock forms holding 60 items, 20 of them moved 10 to 60 mm from their model positions. A published tally gives misplaced items caught and missed, false flags, and each item's position error in mm against the tape, for phone lidar alone and for phone photos with targets.
Open 03 · the brief
Activity dates from site cameras
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A tool that reads the image archive of a fixed site camera and writes down the date each visible activity started and finished, such as each floor's slab, the steel frame, facade panels and the roof. The result is an as-built schedule, the record of what actually happened when. Test it against dates a person reads from the same images and against the contractor's daily reports.
Why it is still open
The December 2025 review of 51 open construction datasets found only 7 that contain video or activity sequences, and called for datasets linked with schedules. This research found camera services that turn site images into time-lapse movies, but no published test of dates read automatically from them. The White Constructions case (see Achieved) shows what a record that does not name the affected work can cost.
Built with what exists
A fixed construction camera's archive of hourly still images; a model that answers fixed yes-or-no questions about images, such as Gemini Decision 1 (see the Note "A thousand questions, one pass"), asked for example "Is the level 5 slab poured?", or an open vision-language model such as Qwen2.5-VL; Python to turn the answers into dates. The camera sees workers, so use an archive whose owner agrees, keep it private, and blur faces in anything published.
First step
Ask a contractor, or a public owner that runs a construction webcam, for six months of hourly stills and the matching daily reports, and write the 40 questions.
Done looks like
One archive of at least six months and 40 activities with start and finish dates read by a person. A published table gives the tool's error in days for each date, its median and largest error, and every date where the images and the daily reports disagree.