As published on Intelligent Building Systems by Zach Mortice..
By nature, construction sites experience rapid change over a relatively short period of time, both in terms of project progress and the ongoing movement of labor, materials, equipment, and money. This dynamism creates an environment challenging for artificial intelligence and machine learning to infiltrate—but when they do, they will face an industry rife for automation.
“As compared to more traditional IT, which is more geared to rule-based, predictable processes like manufacturing, AI has the potential to deliver dynamic digitalization at scale,” says Christoffer Hernæs, chief digital officer at Skanska Norway.
The number of construction firms that are using or plan to use AI is north of 90%, estimates Peak’s Decision Intelligence Maturity Report 2022. Leading-edge firms are embedding AI into their initial project planning and management regimes. As construction sites and buildings are wired with smart infrastructure, the question for early adopters is who—or what—will harness the accompanying torrent of data.
Construction AI is well suited to organize this data and translate it into actionable plans and processes. Though the process of data capture is not adding value, “the insights that you generate from that data is where the value manifests itself,” says Mani Golparvar-Fard, a civil engineering professor at the University of Illinois at Urbana-Champaign (UIUC) who studies AI in construction.
These insights can be reactive or predictive, pinpointing individual building elements that are falling behind schedule; suggesting ways a project can get back on track after a disruption; or identifying which projects have a higher level of safety risks. Much of this intel comes from simple, but time-consuming administrative work—or as Golparvar-Ford calls it, pulling from the tree of “low-hanging fruit: repetitive tasks.”
San Francisco–based DPR Construction robotics lead Henning Roedel agrees. “We’re going to see [teams becoming] more productive," he says, "because they can get the data they need to make decisions much more quickly.”
Second set of eyes
Today, AI is commonly deployed to coordinate construction or monitor progress either from the job trailer or off-site. Reality capture and computer vision, which scans sites to decipher project progress and identify safety concerns, comprise one broad category of AI construction applications. AI is also used to identify time and cost efficiencies in processes such as project sequencing and in the commissioning of smart building components. Rarely is AI directing anything other than a human, though robots that can dynamically react to changing conditions are entering the marketplace.
And while interest in AI is high, adoption remains low. Construction is notoriously a conservative field that often sees technology as extraneous to business. This cautiousness is also driven by construction’s low tolerance for error, huge budgets, long lead times, and high accountability. The wildly popular chatbot applications where AI currently has the most public visibility are notably free of these concerns. “If we were able to provide data that was 90% sure, that [remaining] 10% could be enough for our customers to say, ‘I’m not going to trust this,’” says Mohsan Alvi, the London-based chief science officer at Disperse, a reality-capture AI construction app.
Disperse and other reality capture applications, like OpenSpace, can interpret site activities from photographs and compare progress to supplied schedules and construction drawings. Disperse relies on its teams of humans to visit and photograph construction sites weekly, generating between 8,000 and 20,000 photos per session. It uses AI to compare week-over-week progress and defers to people when ambiguities arise.
Though more time intensive, this deference narrows the margin of error and can offer a faster path to AI self-sufficiency, Alvi says. AI chatbots like ChatGPT work today because of “six years of human training and involvement” during which human users parsed the quality of responses. Construction AI does not have this history. Alvi sees the layer of human review as providing the training feedback.
From its bank of 11 million annotated staff-photos of construction sites, Disperse can now identify more than 680 building components, including structural elements (walls, slabs, and beams), elevators, stairs, fireproofing, façades, and finishings. It can identify 17 different roof elements alone, including drains, gutters, and sheathing.
In what Alvi calls a “construction information model,” Disperse can recognize items of interest and concern to its customers, who are largely general contractors, but also include developers and owners. The app can highlight incorrectly installed components, work that’s been undone, or work happening out of sequence.
In the future, Alvi hopes Disperse will pull richer data from the visual material it collects. This could be measurement calculations via 3D models derived from lidar scans, predictive risk analysis, and ideas for getting a project back on schedule with the help of historical data.
Risk management and decision making
Based in Menlo Park, Calif., ALICE Technologies designed its eponymous construction planning application to also troubleshoot when things go wrong. Alice uses parametric algorithms, similar to architectural design software that synthesizes innumerable formal arrangements of space under user-specified constraints, to generate an array of construction process solutions for contractors. Chief marketing officer Phil Carpenter calls it “construction optioneering.”
At the outset of construction, users upload a building’s 3D model or BIM and define labor and equipment availability, crane locations, and material delivery constraints, like traffic patterns. Weighing labor and material costs, time, and user constraints, ALICE details construction operation options, quantifying the impact of different decisions, such as adding or removing a crane on-site.
ALICE also updates the construction process model as plans change mid-build to recoup time or money in the case of delays or mishaps. Users can model contingencies in advance and identify alternatives when the cheapest, fastest, and most efficient plan falls apart, as it invariably does. “If you plan for something that’s too off the charts, your probability of [success] is not so high, and risk reduction is a smart choice,” Carpenter says.
Understanding how different contingencies can reshape a project helps generate more accurate bids. All told, it adds up to a wildly ambitious goal for ALICE Technologies: increasing the industry’s efficiency to reduce the global cost of construction by 25%.
Skanska is also looking at optimizing the construction process, but with a focus on the movement of machines and equipment on-site. The company is developing an AI platform that optimizes usage of mass haulers, or trucks, to reduce idling downtime, fuel consumption, and maintenance costs. It is developing this technology with Ditio, a Norwegian startup in which Skanska has an equity stake.
Skanska’s initial research found that machines on a typical construction site idle about half their time, most often while waiting in queue. To minimize idling, Hernæs says, Skanska developed algorithms that enable live, dynamic routing by reorganizing circulation routes and staging patterns.
This optimization system, which Hernæs expects will launch to the public this fall, can suggest routing and machine fleet composition day by day, as environmental conditions change. “It’s more of a co-pilot than an autopilot,” he says.
In the future, Skanska envisions this system adapted for the predictive maintenance of equipment and vehicles. By considering hauler utilization rates and cargo types—trucks that haul boulders will need maintenance more often than those that haul soil—“the hypothesis is that not only can we predict when a machine will break down, but also [we’ll] be able to reduce unnecessary repairs,” Hernæs says.
More tried-and-true predictive analytics exist in the construction safety field. Newmetrix, recently acquired by Oracle, analyzes construction management data and integrates reality capture functionality to predict which 20% of projects will experience 80% of safety incidents. Drawing data from construction management platforms, such as Oracle’s Aconex, Newmetrix examines weather, safety procedure observation rates, labor hours, and site personnel to gauge risk. It also incorporates 50 to 75 photos taken weekly by construction crews, making it a subset of reality capture AI platforms. Trained on 17 million construction images collected over time from customers, Newmetrix can detect more than 100 types of on-site safety hazards, including standing water and opportunities for slips, trips, and falls. It flags missing safety gear, messy work sites, and the risks associated with tall buildings, annotating images with safety information.
Newmetrix can also read a construction schedule and adjust its attention on-site accordingly, creating a “dynamically ranked risk list,” says Joshua Kanner, Oracle senior director of products and strategy–Construction Intelligence Cloud.
Insights for smart infrastructure
AI also plays a vital role in the installation and connection of smart building infrastructure and IoT devices. With sensors and systems taking round-the-clock measurements on occupancy, lighting, HVAC, and security, smart buildings generate and collect a mountain of data to make them function as advertised. AI can churn through this information more thoroughly and effectively than any human.
If clients can commit to installing smart building infrastructure a month or so before doors open to the public, then “you can start to use AI to do sanity checks on all the data in real time,” says Buildings IOT CEO Brian Turner. “Now when the commissioning agent shows up, they get a report [of] everything that’s passed and the things that haven’t passed.”
This troubleshooting is superior, he says, to a human spot-checking piles of data for anomalies. For example, a commissioning agent might check that one building section is heated or cooled to the correct temperature at a specific point in time. However, if that temperature is unchanged for six weeks, suggesting a sensor malfunction, the agent will no longer be around to notice.
Buildings IOT’s own building management platform, onPoint, uses AI to troubleshoot and prioritize issues. A building or network that takes 18 to 36 months to complete will reasonably need more than two weeks of commissioning to be fully operational and online, Turner says. “But if you can use machine learning to identify exactly where problem areas are, you can have a better chance of [completing commissioning by the end of construction]. You can probably get rid of 60 to 90 days of post first-day-of-business work effort and have happier clients.”
Individual and industrywide efforts
At UIUC, Golparvar-Fard and his colleagues are collaborating with Carnegie Mellon University to create the Institute for Artificial Intelligence in Construction. Their goal is to establish foundational ideas, technologies, and a workforce that allow AI research to support safe and productive construction and maintenance of civil infrastructure systems. “Many of these topics have been explored for, say, the past decade in academia,” Golparvar-Fard says, “but [rarely] do you see them transform into any real solution that are applied widely in the industry.”
One research area of interest is the integration of whole-language models like ChatGPT in construction AI to produce, for starters, a chatbot that serves as the user interface to any construction AI platform. The industry could use better integration between ambulatory robots and imaging scanners so the former can reliably surmise their exact location relative to surrounding objects and people, Golparvar-Fard says. Currently, AI-directed robots have a limited range of dynamic reactions to site conditions. The Hilti Jaibot can drill into ceilings, change torque when punching through different materials, and self-correct if it drills into rebar. But not much else.
AI-directed robots are “generally good at one thing,” Roedel says. However, he has been impressed with Boston Dynamic’s Spot robot, which became commercially available only recently. DPR is using the caninelike machine to gather reality capture data on job sites, and Spot does an admirable job of avoiding people and obstacles independently. Roedel says on-site exploration—to see “what’s been built, when it’s been built, where it’s been built, and if it’s been put in the right location”—is an ideal role for AI-directed robots.
But single-minded robots are not what’s limiting construction AI. Rather, it is the collection and dissemination of construction data to improve AI algorithms. While amassing hordes of data is useful for the construction AI industry, it’s not particularly useful for its clients. That’s one reason, Roedel says, “we need to get better at automating the data collection.”
Yet, data generated on construction sites can be proprietary and owned by the building owner or contractor. As such, sharing site data and findings industrywide can seem disadvantageous, even though it could advance AI immeasurably. “The more data you offer to the algorithm, the better it learns different scenarios,” Golparvar-Fard says. But “what if the best practices are being learned by some algorithm, and their competitors are tapping into it for the purpose of improving their own process?”
Erasing the data-sharing boundaries between individual, proprietary processes and efforts is the only way to effectively use the vast information produced by construction sites. And this would widen AI’s experience of the world and its role in it for everyone's benefit.