Tech informed by AI has been slow to reach construction sites, but that’s about to change.
Imagine a construction site, shrouded by the dark.
Night works are underway to create an underpass through a busy intersection, and the site is buzzing with activity. A monitor watching live feeds of the work notes a risk: if one of the excavators reverses too quickly, it could hit machinery that’s not necessarily obvious to the driver in the excavator’s cabin. The monitor also logs that one of the cranes is operating at a rate that’s slower than expected, and may be beginning to fall behind schedule. Both of these things are flagged and sent to a supervisor for review.
The catch? The monitor isn’t human: it’s a neural network.
Big construction sites demand millions of fiendishly complicated calculations: the order things need to happen, the way heavy machinery moves and people move around it, and – of course – the cost and time taken to do everything. After years of inertia in the industry, artificial intelligence (AI) is slowly beginning to change the game.
AI that predicts projects
“It’s a new frontier for us,” says Jack Hutchinson, associate director at Hutchinson Builders – one of Australia’s major construction firms.
Hutchinson is currently looking to use two possible AI tools, both predictive: one which estimates the costs of large projects as they progress, and one which predicts the time they’ll take.
“[We’d be] inserting a hoard of historical data about construction programs – so timelines – and having the machine learn from that and predict accurately where they think there will be time blowouts,” explains Hutchinson.
The cost prediction would work in a similar manner. “Essentially, the managing director would have a dashboard of all the projects. And at the start of each project, when we’ve been awarded the contract, we input the data into that cost prediction tool.”
Hutchinson views these tools as “a second set of eyes” for managers in the company.
“For example, if whatever AI software we were using said ‘you’re actually going to be three months late and $2 million over budget’, it’s still up to a project manager or a team leader or managing director or CEO, to make a decision based on that information.”
It’s like a weather forecast, for construction: there’s so much chance this project will cost an additional $50,000, and so much chance it’ll be finished before deadline. Bring an umbrella to the next meeting.
One of the challenges is the historical data used: as all past projects have been predicted and reported by thousands of people, all with different motivations for record keeping, it can be difficult to analyse. Hutchinson is interested to see how the tech companies they’re talking to will clean the data properly.
At the moment, putting AI directly on the construction site is beyond the scope of Hutchinson Builders. “There definitely will be further advancements in technologies on site. But I think a lot of these things are still in their infancy, and they still need to be proven.”
At Monash University, a team of engineers is setting out to prove one of them.
AI that models equipment
Some construction companies now use AI to monitor live feeds of big construction projects. First, cameras are placed on site to capture thousands of images of the equipment in operation. The images are then labelled manually, and fed to a neural network for training. After that, the AI can watch security footage from the site to give the builders extra information they’re looking for – whether it’s about safety, productivity, or something else.
There are a few problems with the system, which is why it’s still not universal among construction companies. For one, the manual labelling is time-consuming and riddled with human error. For another, the AI training requires a lot of footage, which workers aren’t necessarily comfortable with.
“Obviously there are some issues about facial recognition, and identifying individuals,” says Mehrdad Arashpour, an associate professor in construction engineering and management at Monash University.
Arashpour’s team of PhD and postdoctoral students has developed a technology that skips this problem by simulating construction machinery. Using a game design engine, the program creates, randomises and automatically labels three-dimensional models of heavy machinery.
The variables of their models can be jiggled, to reflect the real world. “It can be the randomisation of the equipment itself in terms of pose and texture,” says Arashpour. “It can be the scene randomisation, like the level of lighting – different sunny days, cloudy or night conditions. It can be the floor texture, or background texture, or field of view of the camera.” The options seem to be limitless.
These models are then fed to the neural network, which can analyse what the managers are interested in – for instance, which things are generating safety risks, which things are working slower or faster than expected, and why.
The synthetic images are as accurate as photographs, and much more numerous. “One particular model of the Caterpillar excavator, generated half a million images,” says Arashpour. “It is very quick. In less than two hours, we had the whole data set ready for training.”
And once it’s trained, the AI no longer needs thousands of images for training.
“Time lapse cameras or surveillance cameras that are usually installed on construction sites will be sufficient for the framework now to make inferences,” says Arashpour.
This technology – which is currently under patent review – has been tested on two construction sites, to monitor safety and productivity.
Once patented, here and overseas, Arashpour and his team hope the program could also be used in other industries. “For any setting that you have heavy equipment, this technology is very useful. What immediately comes to mind is construction and building, mining and manufacturing, even agriculture.”
The future of building
Construction and manufacturing are often treated like two sides of the same coin. But their adoption of AI and automation technology shows a key difference: even the simplest, most iterative construction projects will be subject to different environments.
“Construction’s just not the same as manufacturing,” says Hutchinson. “There’s physical context to the site as well. A site in Bondi is different to a site in Brunswick… everything’s tailored to the environment that it’s in.
“Construction has been pretty stagnant from a productivity standpoint for decades, whereas manufacturers continue to get more and more productive, because they’ve been able to automate processes that humans used to do.”
Hutchinson sees the industry beginning to change – for instance, with increasing amounts of offsite fabrication.
“For example, we’ve done module bathrooms, where we construct the whole shell of the bathroom off site, and then crane it down and drop it into a high-rise building,” he says “The benefits are you can treat it more like an assembly line in manufacturing, and you can automate a few more things in the factory environment… I think there’ll be more of that. And it’s been heading that way for a while.”
But the central purpose of construction – building things in a specific place – prevents it from becoming too iterative, and thus, too automated.
This is why Hutchinson doesn’t think the spectre of job replacement is too much of a concern in the construction game as AI is brought into the industry. It’s being used to support decisions, rather than supersede them. “I can’t see that changing anytime soon,” he says. “I’m sure people have said that about many industries, and then they do change… But at this stage, I don’t think the actual delivery of the buildings on site will change too much.
“You can’t produce every architecturally designed high-rise building in a factory. It takes craftsmanship, it takes on-site labour.”
Arashpour agrees. Much of the on-site AI work focuses on safety, which generates more work, not less.
“At the moment, safety managers and safety personnel on site have to monitor operations all the time,” says Arashpour. “They want to add an additional layer for safety or productivity monitoring to what they have. So they are not replacing the manual processes they have at the moment but adding a decision support system.”
Arashpour says making the industry safer is one of his chief motivations. “I think there is much capacity for improvement and growth in construction and infrastructure industries – hopefully by using digital technologies.”
Ellen Phiddian is a science journalist at Cosmos. She has a BSc (Honours) in chemistry and science communication, and an MSc in science communication, both from the Australian National University.