Construction robots and physical AI: what's actually working on jobsites in 2026

Robots are already doing production work on construction sites — just not the humanoid kind. The story that actually moved in 2026 is excavators: three separate companies now sell kits that bolt autonomy onto the dozer or excavator a contractor already owns, and one of them, Bedrock Robotics, says its machines are digging real earthwork on active GC jobsites with nobody in the cab, nobody piloting remotely, and nobody spotting on the ground. What isn't happening yet is a walking, general-purpose humanoid doing varied tasks on an open site — that's still years out, whatever a trade-show demo suggests.

This guide is for the GC ops lead or trade sub owner deciding whether any of this is worth a pilot this year, not the person trying to time when robots take over the trades. It separates what's shipping today from what's still a lab demo, and it names the cost and liability questions vendor pitches tend to skip.

What construction robots are actually deployed on jobsites right now?

Five categories have moved past the demo stage into repeat, paid use on real projects — and a sixth, autonomous heavy-equipment retrofits, went from supervised pilot to unsupervised production in the space of a few weeks this past August.

Robot type	What it does	Example vendor	Where it stands in 2026
Excavator/dozer retrofits	Bolt-on sensor and compute kit that runs excavation, trenching, and loading with reduced or no human oversight	Bedrock Robotics, Gravis Robotics, Komatsu/AIM Intelligent Machines	Fastest-moving category of the year — see below
Layout robots	Prints BIM floor-plan coordinates directly onto the slab for every trade, replacing manual chalk-line layout	Dusty Robotics (FieldPrinter 2)	In production; covers 10,000–15,000 sq ft/day with one operator at up to 1/16" accuracy (Dusty Robotics)
Bricklaying robots	Spreads mortar and lays brick along a laser-guided line for long, straight wall runs	Construction Robotics (SAM100)	In production on commercial masonry jobs; roughly 6x a human mason's daily output (ASME)
Drywall-finishing robots	Tapes, muds, and sands drywall joints to a Level 4 or 5 finish	Canvas (acquired by JLG, Jan. 2026)	In production; cuts finishing schedule roughly in half (ENR)
Demolition robots	Remote-controlled tracked machines that break, cut, and clear structure from up to 984 feet away	Brokk	Mature product category, decades of jobsite use (Brokk)
3D-printed structure	Robotic gantry prints multi-story wall structures layer by layer from concrete	ICON (Titan)	Commercially launched March 2026; first outside-builder systems ship early 2027 (3D Printing Industry)

What all six have in common: they're purpose-built for one repeatable task — a straight wall, a flat slab, a defined excavation plan — not general-purpose robots that navigate an open, changing jobsite the way a person does. That distinction is still the entire story of where this technology stands, even with a no-operator excavator now moving real dirt.

What changed with autonomous excavators in mid-2026?

Three announcements landed within about three weeks of each other, and together they're the clearest signal yet of where construction-robotics capital and technology are actually converging: retrofit kits for the excavators and dozers a contractor already owns, not new autonomous-native machines.

• Late July 2026: Komatsu and its EARTHBRAIN subsidiary struck a strategic partnership with AIM Intelligent Machines, a Seattle-area startup, to bring autonomous operation to bulldozers and hydraulic excavators. Komatsu says the system is already in commercial deployment on U.S. customer sites, with a more cautious Japan launch planned for 2027 (Komatsu; GeekWire).
• August 17, 2026: SoftBank put $200 million into Gravis Robotics — a Zurich, ETH-spinout startup — as the sole investor in what Gravis calls the largest Series A in construction-robotics history, valuing the company at roughly $1 billion. Its Gravis Rack retrofits onto Caterpillar, John Deere, Volvo, and six other equipment brands, and the company claims up to a 30% productivity gain over manual operation — a company-reported figure, not an independently audited one (Bloomberg; Forbes). We covered what the round does and doesn't change for a contractor evaluating retrofit kits here.
• Same day, August 17, 2026: Bedrock Robotics — founded in July 2025 by ex-Waymo engineers and backed by a $270 million Series B led by CapitalG — said its own retrofit kit had moved from supervised testing to full autonomy on live jobsites for Sundt Construction (a Nevada water treatment facility, following an earlier 65,000-cubic-yard supervised-autonomy phase on a 130-acre Phoenix site), Zachry Construction (1.2 million cubic yards of civil sitework), and Champion Site Prep in Texas. No cab operator, no remote pilot, no ground spotter — a site manager sets what to dig and how deep, and the machine executes it, stopping automatically if a person or unrecognized object gets too close (GlobeNewswire mirror; ENR). Full details on what "no operator" changes for a site safety plan are here.

The detail that matters more than any funding number: Sundt and Zachry are large self-performing GCs with heavy civil capability, running this on named infrastructure jobs, not pilot customers testing a demo unit. That's a different signal than a bricklaying robot or a layout printer — this is a category where the operator's seat is now, on at least a handful of live sites, empty.

Are humanoid robots working on construction sites yet?

No. Every humanoid robot shipping in volume right now — Figure AI's Figure 03, now built at a rate of one per hour after a 24x throughput jump earlier in 2026, and Boston Dynamics' Atlas — is headed to an automotive, logistics, or retail environment, not a jobsite (Figure AI). Hyundai's plan to deploy more than 25,000 Atlas units starts in 2028 at its Savannah, Georgia metaplant doing parts sequencing, and even that timeline has a real obstacle: the Hyundai Motor branch of Korea's Metal Workers' Union publicly declared in January 2026 that "not a single robot using new technology" would be allowed into the factory without a labor-management agreement first (Tech Times). Every currently-running humanoid deployment — Figure's units on BMW's Spartanburg line, Agility Robotics' Digit moving totes in GXO and Amazon warehouses — is indoor, flat, and climate-controlled.

Battery and payload physics are most of the reason. Current humanoids run 30 to 90 minutes per charge, and Atlas, at roughly 89 kg, safely lifts only about 11 kg — bipedal balance forces engineers to keep battery mass to around an eighth of body weight, where a wheeled machine can dedicate a third of its mass to power instead (Scientific Reports). Lifting rebar, block, or steel means bigger motors and more battery: weight a two-legged robot can't carry without compromising the balance that makes it a humanoid in the first place. That physics is also why the newest, largest humanoid funding round of the year — XPeng's robotics unit Dogotix raising over $900 million at a $6.3 billion valuation in August 2026 — is aimed at retail showrooms and steel-mill inspection, not a jobsite (Bloomberg). More on what that round does and doesn't mean for a construction fleet is here.

Construction is the inverse of a factory floor on every axis that matters: terrain changes week to week, temporary stairs appear and disappear as floors close out, water and debris are the ambient condition rather than the exception, and there's no permanent operations center monitoring equipment the way a plant has. We covered the specifics of that gap in more detail here. If humanoids reach a jobsite at meaningful scale before 2030, prefab and modular manufacturing — environments that look more like factories than traditional sites — remain the far more likely entry point than a stick-frame or cast-in-place job.

What can't a construction robot do yet, and why?

Three things consistently break the current generation of jobsite robots, autonomous excavators included:

1. Unstructured, changing terrain. Every robot in this guide that's actually in production works from a defined plan on a prepared or mapped surface — a slab, a straight wall course, a graded cut. None of them handle an unmapped, mid-demo floor the way a human crew does without a person setting the plan first.
2. Judgment calls mid-task. A layout robot prints exactly what the BIM model says — it doesn't catch a field conflict the way a super walking the same line would. An autonomous excavator executes the dig plan a site manager sets; it doesn't decide to change that plan when it hits something unexpected.
3. Continuous unsupervised operation across trades. Bedrock's no-operator excavators are the exception that proves the rule — and even there, a person still sets the plan and monitors it, and the published safety control is a proximity stop, not full situational judgment. Every other production example in this guide still has someone operating, monitoring, or immediately supervising the machine.

That's why the realistic near-term model — the one McKinsey's research and the Zacua Ventures construction robotics report both land on — is human-robot teaming, not replacement: machines take the narrow, repetitive, high-exposure slice of a task, and people keep the parts that require reading a changing environment (McKinsey; Zacua Ventures).

How much do construction robots cost, and is the ROI real?

It depends heavily on the category, and the industry's own adoption numbers just flipped direction. A year ago, a BuiltWorlds survey found contractor sentiment toward robotics jumping sharply — positive ratings from 74% in 2024 to more than 95% in 2025 — while actual active use fell from 65% of firms to 46% over the same period; contractors liked the idea more than they were executing on it (Construction Dive). BuiltWorlds' newest benchmarking report, published in August 2026, shows a sharp reversal: 79% of surveyed general and specialty contractors reported using jobsite robotics in 2026, up from 29% a year earlier, and the pilot-only group nearly tripled. Improved accuracy (75%) was the most-cited benefit, ahead of reducing manual effort (63%) and addressing safety concerns (56%) (BuiltWorlds).

BuiltWorlds' own reporting on why pilots historically stalled points to two recurring causes that are still worth guarding against: no one on the project team is assigned to own the machine, so small issues accumulate until it sits parked, and the economics get misaligned when one project absorbs the full pilot cost while the payback shows up organization-wide, months later (BuiltWorlds). Neither of those is a hardware problem, and the contractors quoted in this year's autonomous-excavator news — Sundt, Zachry, Komatsu's own U.S. customers — are the large, self-performing GCs with the staff to assign a named owner. Before signing a robotics vendor contract, assign that owner and write down, in advance, the specific metric — cubic yards moved, bricks per day, finishing days saved — that decides whether the pilot continues past 90 days.

Are construction robots safe, and who's liable if one causes an injury?

There is no construction-specific OSHA standard written for robots. Enforcement currently falls back on general machine guarding rules (1910.212), lockout/tagout requirements, and the catch-all general duty clause (OSHA). NVIDIA's Halos for Robotics — an accredited third-party safety framework whose AI Systems Inspection Lab is the first ANSI National Accreditation Board-accredited program for functional and AI safety in physical AI — is building toward filling that gap, with more than 40 companies involved and Agility Robotics' Digit set to become the first production robot shipping with Halos OS. But that first deployment is a warehouse, not a jobsite (NVIDIA Developer Blog). We broke down what Halos does and doesn't cover for construction specifically here.

The no-operator excavators changed the liability conversation from theoretical to immediate. Every existing site safety plan assumes a competent operator is in the seat, backed by exclusion zones and a spotter when visibility is limited. Bedrock's model swaps that for automated proximity detection running on software the crew didn't write and can't inspect on the fly — with no fenced exclusion zone described in the company's own announcement. Liability itself still follows the same lines as any piece of failed equipment: the GC or sub operating the machine carries general liability exposure tied to how it was deployed and supervised, and the manufacturer carries product liability if an investigation traces the failure to a design or software defect. Before any no-operator or retrofit-autonomy pilot agreement gets signed, get three things in writing: the documented stop-distance behavior and who has authority to shut the machine down, what certification path the vendor is pursuing and through which third-party certifier, and confirmation from your insurance broker that the policy actually covers equipment with no operator of record — nobody has litigated that at scale yet, and no contractor wants their project to be the first. For the fuller risk and insurance picture, see our guide to AI risk and liability in construction.

Will robots replace construction workers?

Not on any timeline currently being pitched, and the labor math argues against it as the right frame entirely. The Associated Builders and Contractors estimates the industry needs to attract 349,000 net new workers in 2026 on top of normal turnover, climbing to 456,000 in 2027, with roughly 41% of the current workforce eligible to retire by 2031 (ABC). That's not a workforce robots are displacing — it's a gap nobody has a plan to fill with people. McKinsey's read is the more useful one for a GC deciding what to do this year: figure out which specific tasks — excavation, repetitive material handling, high-hazard demolition — a robot could take off a crew's plate first, and start building the internal workflow changes now, because the companies that waited to adopt cobots and retrofit autonomy are the ones now playing catch-up on those categories (McKinsey; Construction Dive).

What's the realistic timeline for physical AI on jobsites?

Near term (now through 2027): retrofit-autonomy kits for excavators and dozers spread from a handful of named heavy-civil GCs to a broader customer base, following the same competitive pattern that just played out among Komatsu/AIM, Gravis, and Bedrock within weeks of each other. Layout, bricklaying, drywall finishing, and demolition keep expanding on the same production-proven basis. ICON's Titan platform starts shipping to outside builders in early 2027, extending robotic 3D printing beyond ICON's own projects for the first time.

Even the largest heavy-equipment OEMs are still hedging their timelines: Caterpillar unveiled five autonomous machine lines — excavators, loaders, haul trucks, dozers, compactors — at CES 2026, but gave no commercial availability date for any of them, while its off-board Cat AI Assistant only began rolling out in early 2026 with the in-cab version still in final validation (Interesting Engineering). That gap between a trade-show announcement and a shipping date is exactly the distinction this guide keeps coming back to.

Medium term (2028–2030): humanoid robots reach production scale, but in factories and warehouses, not open jobsites — Hyundai's Atlas plan, complicated by its own plant union, and every other humanoid deployment roadmap are built around that kind of controlled setting.

Longer term (2030+): whether humanoids or human-shaped autonomous equipment reach a traditional stick-frame or cast-in-place jobsite depends entirely on whether someone builds and certifies a safety case for unstructured environments — a harder, slower problem than the hardware itself. Watch pilot announcements from large self-performing GCs on named infrastructure or civil projects, not trade-show demos, as the real signal — that's exactly the pattern Sundt, Zachry, and Komatsu's customers just set with autonomous excavation.

The honest read

The machines paying off today are narrow and, until this August, unglamorous: a robot that prints layout lines, a robot that lays brick in a straight run, a robot that sands drywall to a Level 5 finish. What changed in mid-2026 is that "narrow" now includes an excavator digging real earthwork with no one in the cab — a bigger jump than any humanoid demo this year. None of it looks like the robot on the trade-show floor, and none of it is coming for a super's job. The contractors getting real value aren't the ones chasing the flashiest funding headline — they're the ones running a 90-day pilot with a named owner, a specific number to hit, and a safety plan and insurance sign-off written for equipment that no longer has anyone standing next to it. If you're evaluating any robot vendor this year, ask for the safety certification path, the stop-distance behavior in writing, and the ROI number before the pilot starts, not after. And if the labor shortage is the actual problem you're solving for, the paperwork side of that gap is moving faster than the robotics side — worth fixing first, since it doesn't require a capital purchase or a safety case to start.