Industries / Construction and property / Heavy civil
Surveying and earthmoving
Surveying measures and marks the ground, and earthmoving digs, carries and shapes soil and rock to the levels a design calls for.
6
Milestones achieved, from 1904 to 2013.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
1904
Tracks instead of wheels
Benjamin Holt tested a steam tractor that ran on tracks instead of wheels, to help farmers whose heavy machines sank in soft, muddy soil. The tracks spread the machine's weight over more ground. The National Inventors Hall of Fame calls it "the first practical track-type tractor." Wikipedia dates the test to November 24, 1904, and notes that Richard Hornsby & Sons in England patented a chain track in July 1904.
Lessontwo firms reached tracks within months of each other, so a "first" depends on whether the patent, the prototype or the working product is counted.
Credit
Benjamin Holt, Holt Manufacturing Company, Stockton, California
1932
Rubber tires for earthmovers
R. G. LeTourneau ran a dirt-moving business and built his own scrapers. A scraper is a wheeled steel bowl that slices off a thin layer of soil, carries it and spreads it where fill is needed. The Handbook of Texas says that in 1932 he put rubber tires instead of steel wheels on heavy equipment for the first time, after a customer complained that steel wheels sank in sand. In 1937 he designed the Tournapull, a self-propelled scraper.
Credit
Robert Gilmour LeTourneau
1950
Oil pressure works the bucket
Early excavators dug with winches, chains and steel ropes. Hydraulic excavators move the arm and bucket with cylinders pushed by oil under pressure, which gives finer control. Poclain of France says Pierre Bataille built its first hydraulic excavator in 1950; Wikipedia dates the first Poclain model, the trailer-mounted TU, to 1951.
Credit
Pierre Bataille, Poclain
1953
Light measures the distance
The Geodimeter measured distance with a rapidly switched beam of light sent to a far reflector and back. The AGA Museum says it could "accomplish in hours what had previously taken weeks," and calls it "the first commercially available electro-optical distance meter." Swedish physicist Erik Bergstrand invented it in 1947, and AGA sold it from 1953, though the museum page gives 1950, 1953 and 1956 for the first commercial model in different places.
Credit
Erik Bergstrand; AGA, Sweden
2000
Full GPS accuracy for everyone
Until 2000 the US government deliberately degraded the public GPS signal, a setting called Selective Availability. It was switched off at midnight at the end of May 1, 2000, US Eastern time, at the direction of President Bill Clinton. Afterward, by the US government's figure, basic GPS was accurate to 10 to 20 m or better anywhere in the world.
Credit
US government
2013
Machines follow 3D models
The US Federal Highway Administration (FHWA) promoted 3D design models for road building in its Every Day Counts program for 2013 and 2014. The model data feeds GPS machine control, which guides a dozer blade or an excavator bucket to the design surface, so crews need fewer wooden grade stakes. FHWA called it "a mature technology" in use in numerous states, and gave its own estimates of productivity gains of up to 50% for some operations and survey cost cuts of up to 75%.
Credit
FHWA; equipment and survey firms that sell machine control
Watch for
Excavators with empty cabs
Bedrock Robotics says its excavators work with no operator in the cab on paying jobs, including a water treatment plant with Sundt Construction in Nevada, an earthwork site with Champion Site Prep in Texas, and a 1.2 million cubic yard (about 0.9 million m³) sitework job with Zachry Construction. A machine that gets stuck calls a remote person for help.
Achieved whena customer, insurer or regulator publishes hours worked, volumes dug and incidents for the operator-free machines.
Claimed by
Bedrock Robotics, August 2026, as reported by The Robot Report, which cites no outside check
Caterpillar's autonomous machines
At the CES trade show in Las Vegas, Caterpillar previewed five autonomous construction machines, among them a dozer "capable of grading and leveling terrain on its own," a wheel loader and a haul truck. It also showed a Cat 306 CR mini excavator responding to commands given in plain language. The report gives no performance data and no date when the machines can be bought.
Achieved whena contractor reports one of these machines working without an operator on a paying job, with hours and output measured by the contractor or an outside party.
Claimed by
Caterpillar, January 2026
Open
01
Plan sheets to machine models
A tool that reads a road project's PDF plan sheets and writes the finished ground surface as a LandXML file, the open format that GPS machine control loads. The sheets are the plan view, the profile (the road's up-and-down line) and the typical cross sections. Test it on 300 m stretches of five real projects.
The briefCredit · unclaimed02
Scorecard for model excavators
A 1:14-scale hydraulic model excavator, run by a learned control program, that digs a trench in a sandbox to a set depth. Compare it with a person driving the same model by remote control.
The briefCredit · unclaimed
Open 01 · the brief
Plan sheets to machine models
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A tool that reads a road project's PDF plan sheets and writes the finished ground surface as a LandXML file, the open format that GPS machine control loads. The sheets are the plan view, the profile (the road's up-and-down line) and the typical cross sections. Test it on 300 m stretches of five real projects.
Why it is still open
The dozer is guided by computer, but the model it follows is still built by hand from the plans. A presentation by Tony Lusk and Bobby Steele, hosted on the Indiana state government's website, lists the manual steps in Trimble Business Center ("Alignment, Profile, Corridors, Elevating Linework, and Matching Tie-in's"), warns that converting files "often strips data," and quotes field crews reporting "a bust in your model," a line it credits to "every crew, every time."
Built with what exists
Agency plan sets that also come with the designer's own surface file, to score against (the presentation lists "XML of proposed surface if available"); AI models that read drawings and tables, such as Claude Fable 5.2 (see the Note "Three working devices") or Gemini; Python to write LandXML; CloudCompare to compare surfaces. AI models that read drawings, not only text, are what bring this within reach of a small team.
First step
Find one recent project whose bid documents include both PDF plans and a proposed-surface XML file, and build 100 m of it by hand to set the baseline.
Done looks like
For five projects, the tool's surface compared with the agency's at 100 random points each, with the share of points within 0.10 ft (3 cm) published. The presentation gives the vertical tolerance of GNSS machine control as 0.02 to 0.10 ft.
Open 02 · the brief
Scorecard for model excavators
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A 1:14-scale hydraulic model excavator, run by a learned control program, that digs a trench in a sandbox to a set depth. Compare it with a person driving the same model by remote control.
Why it is still open
Field claims come from the companies themselves, and research is often scored only in simulation. An August 2026 paper reports its results on "an excavation simulation benchmark." A University of Notre Dame lab whose depth camera was degraded by dust in a soil box wrote: "we proceeded using predefined trajectories, highlighting the necessity for future research into robust, real-time 3D mapping methods."
Built with what exists
A commercially sold 1:14 hydraulic model excavator with a microcontroller driving its valves; a depth camera such as an Intel RealSense; Hugging Face's LeRobot library (see the Note "No download") with an Action Chunking Transformer policy trained on about 50 remote-control demonstrations, or a robot model that learns from a few demonstrations, such as NVIDIA GR00T 2 (see the Note "Five demonstrations per task"). Low-cost learning from recorded human demonstrations is the advance that makes this possible.
First step
Buy the model, mark the stations in a sandbox, and record 10 human-driven trenches as the baseline.
Done looks like
20 filmed robot attempts and 20 human attempts, each a 60 cm trench with a 10 cm target depth, measured with a ruler at 10 marked stations. Published: completion rate, mean depth error and worst error, with all the videos.