AI, robotics and emerging technology on site

Artificial intelligence, robots, sensors, prefabrication and digital twins attract more attention than any other construction technology. Some of it is hype; some is already changing how work is done. A construction manager needs to judge new technology on evidence: what problem it solves, how well it works in real site conditions, what it costs, what new risks it brings and what data it needs. This lesson surveys the main emerging technologies, with a closer look at how AI predictions should be evaluated and used. Robotics courses (ROB-430 and ROB-440) cover robot design and safety standards in depth; here the focus is on deployment decisions on construction projects.

What AI is good at, and what it is not

Most AI used in construction today is machine learning: software that learns patterns from historical data and applies them to new cases. Typical applications include classifying images (is this worker wearing a hard hat?), predicting outcomes (is this activity likely to finish late?), extracting information from documents (what are the submittal requirements in this specification?) and generating content (a first draft of meeting minutes, a design option, a schedule fragment). Generative AI tools based on large language models have spread quickly since 2023, especially for drafting, summarizing and searching project documents.

Machine learning works best when there is plenty of relevant, well-labelled data, when the future resembles the past, and when errors are cheap or will be caught. It works poorly when data are scarce or inconsistent (as on many construction projects), when conditions change (a new market, a new building type), or when a confident wrong answer could cause harm. Generative AI can produce fluent, plausible text that is wrong, including invented references and numbers, so outputs must be checked by someone who knows the subject.

Evaluating an AI prediction

Suppose a vendor offers a tool that flags activities likely to be delayed. How should a construction manager judge it? The key is to ask how often its predictions are right, in the specific sense that matters. A confusion matrix compares predictions with what actually happened.

Precision answers: when the tool flags an activity, how often is it really delayed? Recall answers: of the activities that were really delayed, how many did the tool flag? A tool can score high accuracy simply by predicting that nothing will be late if most activities finish on time, so accuracy alone is misleading. The right balance depends on cost: if investigating a flagged activity is cheap and a missed delay is expensive, recall matters more.

import random

random.seed(7)
# Simulate 1,000 activities: 15 % are actually delayed.
actual = [random.random() < 0.15 for _ in range(1000)]
# A model that catches most delays but also raises false alarms.
pred = [(random.random() < 0.70) if a else (random.random() < 0.12) for a in actual]

tp = sum(a and p for a, p in zip(actual, pred))
fp = sum((not a) and p for a, p in zip(actual, pred))
fn = sum(a and (not p) for a, p in zip(actual, pred))
tn = sum((not a) and (not p) for a, p in zip(actual, pred))
print(f"TP {tp}  FP {fp}  FN {fn}  TN {tn}")
print(f"accuracy  {(tp + tn) / 1000:.1%}")
print(f"precision {tp / (tp + fp):.1%}")
print(f"recall    {tp / (tp + fn):.1%}")
always_on_time = sum(not a for a in actual) / 1000
print(f"accuracy of always predicting 'on time': {always_on_time:.1%}")

TP 117  FP 109  FN 51  TN 723
accuracy  84.0%
precision 51.8%
recall    69.6%
accuracy of always predicting 'on time': 83.2%

In this simulation the model's accuracy looks similar to that of a trivial rule that always predicts "on time", yet the model catches most of the real delays, which the trivial rule never does. Precision shows that roughly half of its flags are false alarms. Whether that is useful depends on how much a review costs compared with an unnoticed delay. Asking a vendor for precision and recall, measured on projects like yours, is far more informative than a single accuracy figure.

Other questions matter as much as the numbers. What data was the model trained on, and are those projects like ours? Can it explain why it flagged an activity, so the team can act? Who is accountable if it misses something? What happens to our data once uploaded? A prediction that nobody can interpret or act on adds noise, not value.

Computer vision and safety

Computer vision systems analyse images from fixed site cameras, 360-degree walks or drones. Safety applications include detecting missing personal protective equipment, people inside exclusion zones around equipment, unprotected edges, housekeeping problems and blocked egress. Some systems alert supervisors in real time; others produce weekly trend reports. These tools can extend what supervisors see, but they raise issues of accuracy (false alarms erode trust), privacy (workers being watched constantly) and labour relations (whether images will be used for discipline). Successful deployments involve workers and unions early, focus on hazards rather than individuals, and are transparent about what is recorded and how it is used.

Robots on site

Robots are moving from factories to construction sites, mostly for repetitive, precise or hazardous tasks. Layout robots print floor plans directly onto slabs from the model, using a total station for positioning, replacing chalk lines and tape measures. Drilling robots install overhead anchors from the model. Rebar-tying robots work across bridge decks. Brick-laying and block-laying systems, autonomous compaction equipment and robotic total stations are in use on some projects. Quadruped robots carry scanners and 360-degree cameras through sites on repeatable routes to capture progress. Exoskeletons, worn by workers, reduce strain for overhead work.

Site robots face challenges that factory robots do not: uneven and changing surfaces, dust, weather, clutter, people moving unpredictably and limited connectivity. They also need safe working arrangements: risk assessments, exclusion zones, trained operators and compliance with robot safety standards (ROB-430 and ROB-440 discuss ISO 10218 and related standards). Many site robots are deployed by specialist subcontractors who provide the robot and the operator together.

Prefabrication and modular construction

Prefabrication moves work from the site into a factory: MEP racks and risers, wall panels, bathroom pods, mass timber elements and whole volumetric modules. Factory work offers better quality control, safer conditions, less weather exposure and shorter site schedules. It depends heavily on digital information: models must be coordinated and frozen earlier, tolerances must be tighter, and design decisions must be made before fabrication starts. BIM-to-fabrication workflows send model data directly to cutting, bending and assembly machines.

In Canada, factory-built buildings are commonly certified to CSA A277, which covers the factory quality program, inspection and labelling, with on-site work governed by the provincial building code. Volumetric modular housing has attracted renewed public interest as part of housing-supply strategies.

Sensors, IoT and digital twins

Low-cost wireless sensors, part of what is called the Internet of Things (IoT), report conditions on site: concrete temperature and maturity (helping decide when to strip forms or stress tendons), humidity and moisture in materials, vibration near sensitive neighbours, noise, dust, water leaks, equipment location and worker proximity to hazards. The data feed dashboards (Lesson 9) and alerts.

A digital twin is a digital representation of a physical asset that is connected to it by data and kept in sync over time. During construction, a twin might combine the model, schedule, reality captures and sensor data to show the current state of the site. In operation, a building twin links the as-built model and asset data to the building automation system, maintenance records and occupancy data, so operators can see equipment performance and plan maintenance. The term is used loosely in marketing; a useful test is whether the digital representation is actually updated from the real asset and used to make decisions.

Other emerging technologies

Several other technologies are worth watching. Construction 3D printing of concrete walls and small buildings has moved from demonstrations to some real projects, though code approval and economics vary. Augmented reality overlays the model on the real site through a tablet or headset, helping trades see where hidden services run or check installed work. Autonomous and remote-controlled heavy equipment extend machine control (CON-250). Blockchain-based tools have been proposed for payments and supply-chain records but remain marginal. Low-carbon materials and digital carbon accounting tools are increasingly linked to BIM so embodied carbon can be estimated from model quantities.

Rules for AI at work

Regulation of AI is developing. At the time of writing, Canada has no general federal AI statute in force; the proposed Artificial Intelligence and Data Act died when Parliament was prorogued in January 2025. Privacy law already applies to personal information used in AI systems (Lesson 11), and Ontario now requires employers with 25 or more employees to disclose in public job postings whether AI is used to screen, assess or select applicants. Firms should check current federal and provincial rules before deploying AI that affects workers.

Beyond legal minimums, a firm adopting AI should have a simple policy: which tools are approved, what data may be entered (never confidential owner information into a public tool without permission), who reviews outputs before they are relied on, and how workers are informed. The principle is that AI assists people; accountability stays with people.