Kajima Corporation announced on 5 November that it has begun full-scale deployment, across its construction sites, of autonomous robots running autonomy software built by the California physical-AI company FieldAI. The announcement is a four-page PDF from Kajima’s public relations office, listed on the company’s English news page and carried by FieldAI as a press release the same day.
It is a document with unusually specific numbers in it, and the numbers are the reason to read it.
What Was Announced
Kajima — the release names its president, Masafumi Kiryu, and FieldAI’s chief executive, Ali Agha — says the collaboration began in 2023 and has been refined “through operation on real job sites.” To date, the release says, the robot “has completed over 2,000 autonomous missions across a range of robot types, tasks, and sites.”
The move to full-scale deployment has a specific trigger: “An enterprise-level service agreement reached in 2025 moved the collaboration from individual projects to full-scale deployment.” In roughly a year since that signing, Kajima says it has deployed the robot at more than 10 construction sites in Japan, with deployment continuing “to expand to new sites and applications under the agreement.”
The company is explicit about the distinction it wants readers to draw: “The defining feature of this initiative is that it goes beyond demonstrations under limited conditions: the Robot is in continuous use on active construction sites where conditions change daily.”
The tasks being taken over are named, and they are narrow: “site rounds, visual documentation, and 3D scanning.” The stated purpose is to free engineers to “focus more on advanced technical judgment and safety management,” framed against what Kajima describes as “the shortage of workers in the construction industry” and an effort “to realize workstyle reform.”
What the Robot Actually Is
The architecture is the interesting part, because it is not a robot. Kajima calls the deployed thing “the Robot,” but what Kajima and FieldAI have built is an autonomy unit — cameras, LiDAR and an edge AI device — that is “mounted on any robot” and gives it “the equivalent of human ‘eyes’ and a ‘brain.’”
The edge device does the processing on board rather than in the cloud, which the release ties directly to where construction robots have to work: autonomous operation is claimed “even indoors, in underground spaces where GNSS is unavailable, and in environments where external communication is difficult.” That is a sober engineering constraint rather than a marketing line — a site robot that needs satellite positioning and a reliable radio link is a robot that stops working on the floors where much of the remaining work is.
The brain is FieldAI’s Field Foundation Models. Kajima describes them as combining “data-driven learning with physics-based reasoning and uncertainty quantification,” which it says lets the robot “handle situations it has not previously learned and environments that change daily.” FieldAI’s own site characterises the same model class in near-identical terms, as “a new class of model that combines data-driven AI with physics-based reasoning and uncertainty quantification.” The claim Kajima draws from that is a generality claim: “Unlike conventional robot control technology, which requires systems to be customized for specific sites, it functions as a general-purpose ‘brain’ applicable to a wide variety of robots and tasks.”
The operating loop is described concretely. On first startup the robot explores an unknown area and generates a three-dimensional map of it. After that it makes rounds autonomously, avoiding obstacles, taking photographs and point-cloud measurements and updating the map as the site changes. When it finishes it returns to its charging station and uploads what it collected; the data is consolidated automatically and viewable on the web. A user can also specify a location from a tablet or PC, send the robot there, and watch live video from that point.
One thing the release does not do is name the robot the unit is mounted on. That is a gap in this document, not in the record: FieldAI’s own business blog of 6 February 2026 describes pairing its software with Boston Dynamics’ quadruped Spot on construction sites. Kajima’s text, though, deliberately describes the host as “any robot,” and this publication is not going to put a model number into a sentence the company left open.
Three Use Cases, and What None of Them Does
The release lists three applications, “with additional applications planned”:
Autonomous security patrols. The robot patrols “at any time, including on non-working days, late at night, or early in the morning.” On detecting an anomaly it “issues an alert using flashing high-intensity LEDs and audio,” and a responsible person can view the captured video in real time.
Digitizing the entire site. While making rounds it takes photographs and point-cloud measurements; because “construction status and dimensions can be determined from this data,” Kajima says it streamlines progress management. The PDF’s captions describe a before-and-after pair of automatically captured 360° photographs, labelled before and after interior and MEP installation, alongside what it calls high-resolution colour point-cloud data.
Disaster recovery sites. After earthquakes or heavy rain, Kajima says, deploying the robot “in place of people to safely obtain photos, point cloud data, and real-time video from within the site speeds up the initial response” before site safety has been confirmed. The captioned example is exploration and point-cloud measurement inside a culvert the document describes as difficult to inspect.
Read the three together and a line emerges that is worth stating precisely. The robot drives itself: autonomous mobility with no human intervention is the whole technical achievement here. What it does not do is drive anything else. Nothing in these three use cases operates construction machinery. Patrol, document, scan, and in an emergency go somewhere a person should not.
The Line This Publication Drew in September, and Why Kajima Is Not Where It Applies
Analysis. On 2 September, Caterpillar and FieldAI announced a collaboration, and this publication’s read of it was that the applications named were inspection and situational awareness rather than machine control. We wrote then that the thing to watch would be a scope change: “If a future announcement has a FieldAI model influencing what a machine does rather than recording what it sees, the relationship has changed category.”
On the narrow question, Kajima’s announcement lands on the same side: patrols, photographs, point clouds. Under “Future Plans,” Kajima says it “will also apply this autonomy software to other technologies it is developing, advancing further autonomy in machinery and aiming to realize safe, highly productive construction sites where people and machines work together” — intent, in the future tense, with no product and no date attached.
But the Caterpillar frame does not transfer to this company, and it is worth being precise about why. Kajima already runs driverless earthmoving machinery, and has for years, under its own programme: A4CSEL. The clearest English-language account of it available to this publication is not Kajima’s own — the A4CSEL pages on its site currently return a “This page is currently unavailable” notice — but JTC Corporation’s press release of 17 September 2026, announcing a Singapore pilot with Kajima. Its second footnote states that “A4CSEL enabled three operators based 400km away to manage 14 automated dump trucks, bulldozers, and vibratory rollers, with the machines running continuously for up to 70 hours at one point” during the construction of Naruse Dam in Akita Prefecture.
So this is not a contractor edging toward machine autonomy. It is a contractor that has done machine autonomy at dam scale, now running a second and differently-shaped autonomy programme alongside it.
The difference between the two is the thing to watch, and Kajima’s own release puts it in one sentence without naming what it is comparing against: “Unlike conventional robot control technology, which requires systems to be customized for specific sites, it functions as a general-purpose ‘brain’ applicable to a wide variety of robots and tasks.” A4CSEL is machinery automation built around particular works: JTC says it “has been used on various civil engineering projects in Japan,” the Naruse Dam embankment among them. FieldAI’s pitch is the opposite trade: fewer guarantees per site, no per-site engineering. Kajima has now validated the second shape on perception tasks across more than ten sites, and says it intends to take it toward machinery. Which of the two approaches ends up driving Kajima’s excavators is a genuinely open question, and nothing in this release answers it.
The JTC pilot is a hint at how the company hedges. There, per JTC, the autonomous compactor runs A4CSEL — its first deployment outside Japan, the release says — while the autonomous excavator uses a retrofit kit from Gravis Robotics, the Zurich firm SoftBank put $200 million into in August. One contractor, three autonomy stacks, each on a different machine or task.
Where This Sits Among the Japanese Contractors
Kajima is not a pilot customer of convenience. It is one of Japan’s largest general contractors and designers, and this publication has noted before that a Kajima deployment is a demanding signal precisely because the firm does not adopt experimental software for its own sake.
It is also not the only large Japanese contractor in this autumn’s robotics coverage, and the bets differ in shape. Shimizu put money into Noble Machines and its Moby humanoid — a new body, built for the site. Kajima’s FieldAI bet is the inverse: keep whatever body you like, standardise the brain. HD Hyundai’s work with XiteSolution in Korea is a third approach, going at the excavator directly from the equipment-maker side.
Against the wider market the Kajima figures are modest, and that is what makes them useful. Bedrock’s retrofit kits are already doing paid work on named contractors’ sites with autonomy on the machine itself, and this year’s construction-robotics funding is concentrated in a handful of companies. What Kajima is reporting is not a bigger bet but a counted one: completed missions on sites where the floor plan changed while the robot was working.
Groundworks 2026, in London on 11 November, is organised around precisely the gap Kajima claims to have crossed — demonstrations into routine use. Kajima’s release is a data point for that argument, and it arrives six days before the summit.
What the Release Does Not Say
Scoped to this document: it gives no cost, no headcount effect, no number of robots, no named site, and no measure of how much engineer time the robot has actually displaced. The 2,000-mission figure is a count of missions, not a result. The “more than 10 sites” figure is not set against how many sites Kajima runs.
There is also no safety or incident reporting in it, which is a live question for exactly this class of deployment: NIOSH’s bulletin on construction robotics made the point that the factory playbook for keeping people away from machines does not transfer to a site where the machine has to work among them. A robot doing patrol rounds is the mildest version of that problem. It is still a version of it.
What the release does establish, on the company’s own account, is a collaboration dating to 2023, a commercial agreement that changed its shape, a mission count and a site count — which is more than many announcements in this category carry.
