Construction is famously labor intensive: direct labor makes up close to 50% of the cost of constructing a new single-family home in the US, compared to around 6 to 8% of the cost of manufacturing a car. Unsurprisingly, there’s long been interest in automating the construction process to reduce the amount of labor required, and therefore the cost of housing. The earliest attempt at automating construction I’m aware of dates to 1853, with a plastering machine invented by Isaac Hussey, and since then there have been numerous other attempts.
I wanted to get a better sense of what the arc of construction automation looks like over time: what things had been tried when? What worked, and what fell by the wayside? To do this, I used AI to trawl through thousands of old issues of digitized construction trade journals going back to the 19th century, looking for any example of a construction task being automated.
My main takeaway is that successful construction automation basically falls into two categories. One is construction-related tasks that can be done off-site in a factory: this includes things like steel fabrication, precast concrete, window and door manufacturing, and so on. The other category is on-site tasks that are, or can be made to be, extremely factory-like: they can be done by a machine moving along a simple path, doing some repetitive motion over and over again. Historically, almost anything outside these two categories hasn’t succeeded.
Method
I started by downloading every back issue of Architectural Forum (950 issues) and House & Home (370 issues) on USModernist, and every back issue of Engineering News-Record (4,300+ issues) available on archive.org. These are trade journals that cover various aspects of the construction industry, including construction practices and methods, and cover the period from the mid-19th century up to the early 1980s. I then had Claude (Opus 5 Max) search through these issues and flag any instance of some construction task being automated. In addition to this, I also had Claude do a more general web search for examples of construction automation, which included things like looking through the archives of the International Association for Automation and Robotics in Construction (IAARC).
The purpose of this wasn’t for data analysis: I wasn’t counting the number of cases by year or by category, or anything like that. This was purely an existence survey so I could see what sorts of things had been tried and were promising enough that they made it into a trade journal.
As usual with a task like this, the results will be a function of what definitions you choose to operationalize the terms in your prompt. By “automation,” I mean something like “a machine does some substantial portion of a task without requiring human direction.” So this is distinct from “mechanization” or “industrialization,” where a machine is used to reduce the labor required to do some task, but that machine might still require continuous direction or direct operation from a human operator. So something like a power tool or a crane wouldn’t count as automation, because in both cases each individual movement/action of the machine is orchestrated by the human as the work is being done. A bricklaying machine that automatically placed individual bricks without a human operator directing it brick by brick, on the other hand, would count as automation.
Like most definitions, this one gets fuzzy around the edges. In practice most automated systems require some degree of human oversight while they’re running, and whether that oversight counts as “directing individual movements” can be something of a judgment call. So I’d consider a road-paving machine that automatically lays down asphalt “automation,” even with a human driver monitoring the machine continuously. But something like a ride-on concrete finishing machine seems much more borderline, since that seems to require a lot more human direction on a second-by-second basis.
Factory methods
If some construction task can be done outside an actual construction site, in a factory or factory-like environment, there’s a very good chance that, like any factory task, it can at least be partly automated.
An early example of this is concrete batch plants. Through most of the 19th century, concrete was mixed manually by workers with shovels, but by the late 19th century concrete mixing started to be done by machine, and during the first half of the 20th century these became more and more automated, with things like automatic cement weighing, automatic water dispensing, and so on. On the Hoover Dam, for instance, concrete batching was “completely automated.”
Steel fabrication is similar. Steel sections are, of course, a factory-produced product and always have been, and it’s not much of a leap to turn raw steel elements into building components at the factory and to partly automate those tasks. Machines to do tasks like automatically drilling holes in steel beams, automatically welding pipe seams, and automatically painting steel sections started to appear in the late 19th century and remain widely used today.
More generally, essentially any building material — steel, brick, concrete block, plywood, drywall, shingles — will be produced in large volumes in a factory and thus various portions of the production process will be automated.
Concrete, as we’ve noted, started out as a site-produced material but eventually had its production moved off-site. We see the same trajectory with other building components. Windows and doors, for instance, used to be fabricated on-site, but those are now produced in off-site factories with many of the assembly operations automated. Floor and roof trusses were similarly originally site-produced components but are now mostly made in off-site factories with varying degrees of automation. A typical truss factory will automate the cutting of the various wood sections and the pressing of the truss plates, and some truss factories have even higher levels of automation.
Precast concrete is another factory-produced building component: large concrete elements like beams, walls, and columns are manufactured in an off-site plant, then trucked to the jobsite and assembled there. Precast manufacturing doesn’t necessarily have a great deal of automation, particularly in the US, but it can be highly automated.
This is, of course, part of the reason there’s been so much historical interest in prefabricated construction: once you move the actual construction process into a factory, you can potentially automate large fractions of it, achieving (it’s hoped) labor and cost savings. While sometimes prefabricated construction is merely conventional, labor-intensive construction done inside a warehouse, other times the process is highly automated. Levitt and Sons built a highly automated prefab plant as part of Operation Breakthrough in the 1970s, and today companies such as Autovol in the US and Gropyus in Germany run partly automated plants.
On-site continuous processes
The other category of successful automation is on-site tasks that can be made factory-like: a simple, repetitive motion, done over and over again as a machine moves along a simple path.
A lot of the examples in this category involve concrete or asphalt. Concrete is a good candidate for this sort of automation because it’s a thick, viscous fluid; it can be squeezed into the shape of some particular container and then maintain that shape when the container is removed. So you have things like automatic concrete slipforming machines, which extrude a layer of concrete as they advance forward. Oriented horizontally, these can be used to make things like curbs, trenches, and sidewalks. Oriented vertically, these can be used to make vertical concrete structures (though I believe vertical operation is somewhat less automated).
We also see this sort of “automatic extrusion machine” for metal. Gutter-making machines are basically portable roll-forming machines: they take in sheet metal from a coil and shape it into gutters. Similar portable machines exist for metal roof deck and metal siding.
You also see this sort of automation used successfully in the construction of horizontal, continuous infrastructure. Roads, for instance, are paved with a series of specialized machines called a paving train. Working together, the machines work in sequence to rip up and roughen the existing surface of the road and then lay down a new layer of asphalt.
Parts of railway construction can similarly be done by a series of automated machines. Maintenance tasks — things like replacing worn-out railroad ties and tamping ballast — in particular seem to be highly automated. Since the rail is already in place, this machinery can be rail-mounted, performing its various specialized tasks as it moves along the track.
Tunneling is similar. Historically tunnels were done by hand, but around the turn of the 20th century tunneling began to be done by self-propelled boring machines.
Everything else
From what I can tell, up through the early 2000s (and even into the 2010s), almost everything that didn’t fall into the two categories (tasks that can be done in an off-site factory, and tasks that can be done in a factory-like way on-site) didn’t succeed.
Automatic bricklaying machines are the classic example of unsuccessful automation here. As I noted in a previous essay, these sorts of machines have been tried over and over again since the early 20th century, and historically none of them succeeded.
Automatic setting of bricks or concrete blocks seems like it would be the perfect construction task to automate: it’s both physically strenuous and extremely repetitive. But in practice, the task isn’t quite simple and repetitive enough to build a machine to easily do it. Spreading a layer of mortar, and then placing a brick on top of it, while ensuring that both the individual brick and overall wall remain straight and level turns out to be hard to do with a uniform, repetitive mechanical motion: in practice the bricks often end up slightly askew, requiring small, individual adjustments to the bricks to make sure that the wall is being built level. Making nice, smooth mortar joints similarly isn’t easy for a machine to do. SAM, the Semi-Automated Mason manufactured by Construction Robotics, required masons to follow behind the robot to finish the joints and occasionally level bricks that were set askew. Fastbrick Robotics’ Hadrian block-laying robot avoided these problems entirely by using a custom adhesive that doesn’t require spreading a layer of mortar.
Brick walls also often aren’t repetitive enough in the sense that they rarely consist of long, uninterrupted runs of uniformly placed brick. Walls are often short, they have window and door openings, they have joints, they have corners and jogs, and various other deviations from uniformity. A simple machine doing a purely repetitive motion, or even a robot with a greater degree of flexibility, can’t easily handle these sorts of deviations, and historically bricklaying machines needed to be manually readjusted by a person whenever they encountered one of these situations. The SAM bricklaying robot really only worked well on very long, uniform stretches of wall; because it couldn’t turn corners or even set bricks at the end of walls, it needed to be set up fresh at every new wall, which took time and effort.
This sort of difficulty is very common with construction automation. Often there’s some repetitive task that can be automated, but it’s not quite repetitive enough to make it worth setting up some machine to do it. If a person has to manually reset a machine every few minutes and move it to a new area, that’s often not any easier than simply having a person do the task. Builders tried automatic welding machines on jobsites as early as the 1930s, for instance, but you don’t see many such machines used today: buildings rarely have long, uninterrupted continuous stretches of welding that would make such a machine worth it.
Similarly, tying rebar — wrapping a thin wire around two pieces of rebar to tie them together — is simple enough to be automated, but historically it hasn’t been quite repetitive enough to field a dedicated machine for it. (In my experience, it’s often not repetitive enough to even give workers power tools that can do this.) Even today the calculus appears pretty unfavorable: Advanced Construction Robotics started offering the TyBOT in 2020, but you only ever see it demonstrated on concrete bridge decks, which have very large expanses of rebar to tie. (It also has a rebar-placing robot, IronBOT, but not only does this also require large expanses of rebar to make sense — its website states that “[s]etup requires planning and coordination due to the IronBOT’s size and capabilities” — but the actual automation is fairly thin, requiring a human to constantly feed it individual bars.)
The lack of sufficient repetition, combined with difficulty in making a construction site “factory-like,” is what doomed Japan’s push to develop construction robots at the end of the 20th century. Starting in the late 1970s, Japan undertook a major effort to develop all manner of construction robots — column-welding robots, fireproofing-installation robots, wall-painting robots, and so on. By the 1990s, over 100 different types of construction robots had been developed. As I noted in a previous essay, while these systems often worked, they generally weren’t particularly productive:
Though the robots could often perform their tasks extremely efficiently, they required a great deal of setup and teardown time, and often encountered difficulties in buildings that hadn’t been specifically designed to accommodate them (such as narrow spaces that the robots couldn’t reach). They also had trouble dealing with the somewhat unpredictable nature of a conventional construction site — their need to work uninterrupted, combined with safety limits on how closely workers could approach, tended to constrain the rest of the construction process. And even when tasks could be completed faster, it often meant that the bottleneck just shifted to some other part of the process…
To try and overcome these issues, Japan’s largest contractors tried to make jobsites as factory-like as possible. They created huge “skyscraper factories,” huge climbing platforms that would build a skyscraper floor by floor with the aid of all manner of construction robots. But this failed to solve the problem:
The long factory setup time made them impractical for smaller buildings, and even on buildings that were large enough, extensive upfront design and coordination was required to design the building for robotic assembly. In some ways, the move from single-task construction robots to fully automated construction sites just shifted where the difficulties occurred — instead of a long setup (and takedown) time for each individual task, now you had one enormous and costly setup over the entire first portion of the project.
Future
Historically, construction automation has been successful on certain types of tasks: tasks that can be moved off-site into a factory, or tasks that are simple and repetitive enough that they can be done in a factory-like manner on-site. Automation outside of these areas generally hasn’t succeeded. Often this is because the task isn’t quite simple and repetitive enough: while you can build a machine to do it, the time and effort it takes to get the machine set up and doing the work, and orienting the jobsite to accommodate it, obviates any advantage you get from the automation.
I think, however, that we’re on the cusp of this changing.
This historical pattern we’ve observed is, in large part, a technological limitation. For most of history machines have had, at best, very narrow information-processing abilities: they were limited in what they could sense about their environment, and the extent to which they could modify their actions based on that information. Often they simply did the exact same motion over and over again regardless of the circumstances. When they could modify their behavior based on environmental feedback, it was in very limited ways. Without robust information-processing abilities, automation is limited to being done in a factory (where you can break apart your task into a series of small tasks, each one done by some particular machine) or to processes that are naturally factory-like.
The history of welding automation provides an example. The earliest welding machines were fixed automation: they performed some specific welding motion — such as welding the seams on pipes — over and over again. Starting in the 1960s, welding began to be done with robots: now instead of needing to build a new machine, you could simply reprogram the robot, feeding it a different sequence of actions to take. But that sequence was still dependent on a human programmer to determine.
Over the next several decades robotic welding gradually accumulated information-processing abilities: by the 1980s a welding robot had a limited ability to determine its position with respect to the weld and modify its actions accordingly. But the scope of this ability was very narrow: a welding robot could do things like measuring the electrical conductivity through the weld to determine its position, but not anything more complex:
They couldn’t compensate for things like a clamp that impeded the path of a robot, or a part that was too dirty to weld properly, or a part that was oriented vertically instead of horizontally, or any of the other thousand things that a manual welder could automatically adjust for.
Thus even automated robotic welding required a fairly repetitive set of tasks to be worth the time and effort it took to set up.
But this technological limitation is changing, and we’re getting machines that have far more robust information-processing capabilities. Modern automation is increasingly capable of being given a set of instructions and figuring out what actions it needs to take, step by step, to achieve them. This is, in large part, due to modern AI technology: if you feed some AI model a huge amount of training data, you can get something that responds effectively in different situations — a self-driving car that navigates a route and a traffic pattern that it’s never observed before, an LLM that answers a question that it’s never seen before, and so on.
This changing constraint is reflected among the emerging set of construction automation startups. There’s a new crop of construction automation startups that are leveraging modern AI technology and are getting traction and raising large amounts of funding. Roughly half of the VC funding for construction robot startups has gone to companies founded post-2022, and most other large funding rounds for construction robotics have been post-2020. (And this doesn’t include the enormous amounts of funding that have gone into more general-purpose, non-construction-specific robots, like Figure.) It’s increasingly becoming possible to automate construction tasks that were previously resistant to it, and I only see this trend continuing.
It’s not clear to me exactly what form this will take: will we see an army of Optimus robots on construction sites capable of a broad array of tasks, or will it be a bunch of special-purpose robots like Dusty and Canvas and Rufus? I suspect that regardless of how it develops, the future of construction automation does not look like the past.
Great article!
I love how you start with a huge dataset, sorting each innovation useful categories, and then following up with interesting examples with photos.
There's so much external construction to just building the house that drives up cost and drags on efficiency: sewers, electric, water mains, grading, roads/driveways, sidewalks, curbs permitting, etc. It's pretty easy to prebuild framing sections and raise them on-site (or prebuild the whole house and ship it but I hate how those oversize haulers tie up traffic and are you really saving much by having to ship something that way?). The only thing that comes to mind are those screw-in pilings: I've seen them used for large garages and other outbuildings...could these replace the typical concrete foundation (aka pier and beam without the perimeter concrete stem wall) and increase speed of construction?