Industries / Industrial equipment and services / Heavy equipment manufacturing
Controls and automation
Building the electronics and software that report a machine's condition, steer it to a plan, and let it work with no one in the seat.
4
Milestones achieved, from 2001 to 2019.
2
To watch for: announced, not yet confirmed outside.
2
Open deliverables. Nobody has shipped them.
The road
AchievedWatch forOpen
Achieved
2001
Machines report from the field
Komatsu developed KOMTRAX, its machine tracking system, in 1998 and began fitting it as a standard feature in July 2001. A unit on the machine sends its location, hours, fuel use, idle time, warning codes and maintenance reminders by satellite to a data center, where owners read them online. This is called telematics. Komatsu says KOMTRAX is standard on most of its Tier 3 construction machines and free for the life of the machine.
Credit
Komatsu
2008
Haul trucks drive themselves
Komatsu launched its Autonomous Haulage System, which drives mining trucks with no one aboard, in January 2008; its own history calls it the world's first. The mining company Rio Tinto began trials of trucks with the system in 2008. By January 2018 its driverless Komatsu trucks had moved more than 1 billion tonnes at five iron ore mines in the Pilbara region of Australia. In 2017 they moved about a quarter of the material there, each worked about 700 hours more than a staffed truck, and load-and-haul unit costs were about 15% lower, by Rio Tinto's figures. Caterpillar says its own autonomous mining fleet had moved more than 11 billion tonnes by early 2026.
Credit
Komatsu, with Rio Tinto
2016
Tractors steer by satellite
Auto-steer and guidance systems use satellite positioning to track and steer a tractor, sprayer or harvester along planned lines in a field. The US Department of Agriculture's Economic Research Service found them in use on 5.3% of planted US corn acres in 2001 and 58% in 2016, and on 72.9% of sorghum acres and 64.5% of cotton acres in 2019. In 2019, satellite positioning was used on 40% of all US farm and ranch land.
Credit
farm machinery and guidance makers; adoption measured by Jonathan McFadden, Eric Njuki and Terry Griffin, USDA Economic Research Service (2023)
2019
Driverless trucks collide
In February 2019 a driverless truck at Fortescue's Christmas Creek iron ore mine in Western Australia reversed at slow speed into a parked driverless truck. A dropout in the site's Wi-Fi coverage, which carries data to the control center, was believed to have contributed; Fortescue said the autonomous system had not failed. In March 2019, at BHP's Jimblebar mine, an empty driverless truck lost traction in a sudden storm and slid into the path of a loaded one traveling at 27 km/h. BHP had not applied its speed-restriction procedure, and the state regulator issued an improvement notice. No one was hurt in either collision.
LessonA driverless truck depends on the site's radio network and weather procedures as well as on its own sensors.
Credit
a failure; reported by the Australian Mining Safety Journal from statements by the companies and Western Australia's Department of Mines, Industry Regulation and Safety
Watch for
Driverless tillage on farms
John Deere sells a retrofit kit, the Autonomy Precision Upgrade, that adds 16 cameras so that a Deere tractor can make tillage passes with no one aboard while the farmer watches and adjusts it from a phone app. In autumn 2025 it worked only on corn and soybean fields in the US Midwest and only with certain Deere tillage tools.
Achieved whenan independent party, such as a university extension service, reports acres covered and the number of stops and interventions per hour over a full season, on farms Deere did not choose.
Claimed by
John Deere (Michael Porter, go-to-market manager for large tractors and tillage), October 2025
Cat's autonomy leaves the mine
At the CES trade show in January 2026, Caterpillar previewed autonomous versions of five kinds of construction machine: excavators, loaders, haul trucks, dozers and compactors. In September 2026 Brandon Hootman, its vice president for physical AI platforms and construction autonomy, said Caterpillar had hauled more than 3.5 million tons autonomously in an 18-month pilot, and that it and the quarry company Luck Stone were expanding autonomous hauling to two more sites.
Achieved whenLuck Stone or another customer publishes tonnage, uptime and incident figures, or the construction machines go on sale to customers outside the pilot.
Claimed by
Caterpillar, January and September 2026
Open
01
Hobby excavator digs to grade
A model excavator, about 1:14 scale, that learns from recorded human demonstrations to dig a trench to a set depth and slope in a sandbox, with a published test that anyone can repeat. Grade is the designed height of the finished ground.
The briefCredit · unclaimed02
Open test for people detection
A repeatable test, run on a parked machine with the engine off, that places a person-sized dummy at marked points around an excavator or loader and records whether each camera-based people-detection system raises its alarm. The result is a published map of where each system sees and misses people, including one system built from open parts.
The briefCredit · unclaimed
Open 01 · the brief
Hobby excavator digs to grade
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A model excavator, about 1:14 scale, that learns from recorded human demonstrations to dig a trench to a set depth and slope in a sandbox, with a published test that anyone can repeat. Grade is the designed height of the finished ground.
Why it is still open
Laboratories have automated digging on scale models, for example trajectory control on a scale hydraulic excavator reported to the Japan Society of Mechanical Engineers (Ito and colleagues, 2020), and the AI-MaSi project runs a path planner developed in NVIDIA Isaac Sim on a modified radio-controlled excavator. A hobby project, ALSET, set out to train a toy excavator by demonstration and has been on hold since 2022. In September 2026 Shuai Zhao and colleagues reported a scaled hydraulic excavator that learned digging from expert demonstrations; it measured payload per scoop (6.52 kg against 2.68 kg for a fixed motion), not depth to a target. Searches of GitHub, Crossref and the web in November 2026 found no published dig-to-grade test on an off-the-shelf model, with a pass rate and depth errors that others could reproduce.
Built with what exists
The open-source 3D-printed RC excavator published by ProfBoots on GitHub, or a hobby hydraulic model, with angle sensors added to the boom, arm and bucket; LeRobot to record demonstrations and train a policy (see the Note "No download"); an open robot model such as Runway's Praxis-1 (see the Note "Open robot model weights") or NVIDIA's GR00T 2 (see the Note "Five demonstrations per task"); a depth camera such as the Luxonis OAK-D to measure the trench.
First step
Build or buy the model, add the angle sensors, and record 50 trenches dug by remote control with LeRobot.
Done looks like
20 filmed trenches, each 30 cm long at a set depth of 5 cm, with depth measured at 10 points along each. Published: the share of trenches with every point within 5 mm of the set depth, the mean error, and the same figures for a person driving the same model by remote control.
Open 02 · the brief
Open test for people detection
The first to publish is credited on this line, and the deliverable moves from Open to Achieved.
What to build
A repeatable test, run on a parked machine with the engine off, that places a person-sized dummy at marked points around an excavator or loader and records whether each camera-based people-detection system raises its alarm. The result is a published map of where each system sees and misses people, including one system built from open parts.
Why it is still open
Contractors fit AI cameras to keep machines from striking people, but published field results count alerts, not misses. Fletcher Construction in New Zealand trialed three products on a crane, a skid steer and an excavator for two months in 2023 and reported more than 1,500 zone breaches and a 60% fall in alerts over time, with no detection rate. A British and international standard sets performance requirements and tests for object detection systems on earth-moving machinery, but searches in November 2026 found no published side-by-side results for commercial systems.
Built with what exists
A mannequin or foam dummy; a measured grid marked on the ground; the systems under test; for the open system, Ultralytics YOLO11 running on an NVIDIA Jetson Orin Nano with three or four cameras.
First step
With a rental yard's permission, mark a grid around one parked excavator and photograph the dummy at every point from the machine's camera positions.
Done looks like
Each system tested at 100 grid points around the same machine, by day and at night, with the dummy standing and crouching, and the miss rate for each zone published with the photos and the test procedure.