Earlier this month, a robot called Douglas showed up to work on a Tilbury Douglas construction site in Britain. He weighs around 32 kilograms, was built by a Chinese company called Unitree, and costs roughly the same as a small van. Tilbury Douglas gave him a brain: an NVIDIA AI computer mounted on his back, with the Unitree handling locomotion and the NVIDIA handling perception, spatial mapping, and object recognition.

For the last ten weeks, he's been doing something no humanoid robot had managed before on a UK site: turning up every day to do an actual job. Mark Buckle, Tilbury Douglas's Technical Director, has been the man behind the deployment. The way he describes what Douglas is doing (and what he is not doing) is probably the most honest account of where this technology actually sits right now.

The labour picture in construction hasn't improved, and by most measures it's getting worse. We've written about contractors walking away from work as the skilled trades pipeline thins out. The UK's Construction Industry Training Board puts the shortfall at 251,000 extra workers needed by 2028, and around 24% of the current workforce is already over 55. Half the people who know how to do this job are within a decade of retirement, and the pipeline behind them isn't close to filling the gap.

Capital is moving fast in response. Morgan Stanley projects the broader humanoid market could reach $5 trillion by 2050, describing it as potentially twice the size of the auto industry. The money is real. But for construction, it isn't flowing here yet. It's moving into automotive plants, logistics warehouses, and electronics assembly lines: environments that already work at scale. Construction is watching that happen from the outside.

The headline deployments are in automotive manufacturing. Figure AI's humanoid supported production of more than 30,000 vehicles at BMW's Spartanburg plant, logging over 1,250 hours. Agility Robotics' Digit is running RAV4 logistics at Toyota Motor Manufacturing Canada. Both are structured, predictable, high-repetition environments: fixed lighting, known layouts, standardized workflows. A live construction site offers almost none of those conditions.

Humanoid robots remain expensive systems, with most advanced commercial and research-grade units typically priced well into the six-figure range. That is what makes Unitree’s pricing notable: the company’s Unitree G1 humanoid starts at $13,500, far below the cost traditionally associated with bipedal humanoid platforms and closer to the economics required for real-world deployment at scale (Unitree Robotics). Douglas too sits in this price range, thus expanding the market for affordable humanoids.

Adam Jonas, head of Global Autos and Shared Mobility Research at Morgan Stanley, put the timeline plainly: "Adoption should be relatively slow until the mid-2030s, accelerating in the late 2030s and 2040s."

A BMW production aisle and a construction site have almost nothing in common from a robot's perspective. Layouts change daily, sometimes within a shift. Multiple trades work in overlapping spaces. Ground conditions, lighting, and network connectivity vary by the hour. Humanoids navigate by processing continuous data about their surroundings: sensor feeds, spatial mapping, and contextual inputs. On a live site, that data environment is adversarial.

The specific technical gaps that matter most for construction aren't solved problems:

Insurance liability for on-site humanoid incidents is also unresolved. Mark is candid about this: Tilbury Douglas is working through it with insurers in real time, finding a new risk to document most days. McKinsey's construction robotics report puts large-scale deployment in construction at a decade away.

Mark's team didn't start with the robot. They started with a problem: site engineers were spending one to two hours a day on walk-arounds, capturing progress photos, overlaying imagery against the BIM model, and logging health and safety observations. Repetitive work that absorbed skilled time but demanded very little of the judgment those engineers were hired for. They identified three tasks, then worked the technology backwards to fit them.

Douglas's job covers programme tracking (visual progress updates), quality checks (overlaying LiDAR capture against the BIM model to flag discrepancies), and H&S observation (identifying trailing leads and hazards during his walk-around). He doesn't work at night and doesn't work unsupervised. As Mark puts it: "just take those repetitive tasks away and let us focus on being amazing builders."

The workforce reaction has been more curious than resistant. Mark reaches for an analogy that lands better than most:

Most of the fears people bring to Douglas, as Mark notes, are about capabilities the robot doesn't have and isn't being asked to develop. The skills of a seasoned bricklayer aren't at risk here. The clipboard-and-camera hours of a junior site engineer are a different story.

McKinsey puts large-scale construction humanoid deployment a decade away. That's not a reason to wait. The gap is a preparation window, and as we've explored in the context of AI strategy more broadly, the firms that benefit most from new technology are the ones that had laid the groundwork before arrival. Mark's framing is deliberately narrow: of the ten or so repetitive tasks currently absorbing 30 to 40% of your site team's time, pick one and solve it. Not everything at once.

Three areas that are worth moving on to now:

There's also a recruiting angle. Mark raises that the industry hasn't fully sat with it. Construction's image problem is real. A firm deploying robots on a BIM-integrated site sends a different message to students and graduates than "dusty and dirty." It's a secondary benefit of the Douglas program, but probably an underrated one.

Douglas is one data point. The methodology behind it (start with the problem, work the technology backwards, scope narrowly) is the transferable part. And as real-world proof goes, it's more useful than anything currently doing backflips in a demo reel.

POWERED BY: