Do Driverless Mine Trucks Win by Being Steady, Not Fast?

Aslower truck that moves more rock sounds like a riddle. Our research desk says part of the answer may lie in a 1960s queueing formula, and the same logic sits under the AI systems now running medicine and logistics. The lesson for a careful reader is about consistency versus raw speed.

An open-pit mine is a loop. A shovel loads a truck, the truck hauls rock to a crusher, dumps, and drives back. In queueing terms the shovel is a server and the trucks are customers that keep rejoining the line. This is a closed cyclic queue.

Since roughly 2008, large mining producers have replaced drivers with computers and reported productivity gains. The first commercial use was at Codelco's Gaby copper mine in January 2008, followed by Rio Tinto's West Angelas iron-ore mine in late 2008. Suncor began an oil-sands pilot in 2013. The usual explanation is hours: Rio Tinto reported that each autonomous truck ran about 700 more hours a year than a staffed truck. Yet the autonomous trucks also drive slower, braking early and lining up carefully, and a UBC thesis on autonomous haulage found they still had shorter cycle times. A slower machine that moves more rock needs an explanation.

The desk's proposal: a meaningful part of the gain comes from lower variability in cycle time, not speed and not only extra hours. Human drivers differ lap to lap. One brakes late, one dawdles, one slow driver holds up others on a road with no overtaking. A computer drives the same lap every time. Steadier arrivals mean the shovel idles less and queues less.

The queueing math is real and textbook. John Kingman's 1961 approximation for the average wait in a busy single-server queue is the product of how close the server is to full use, how irregular the arrivals and service times are, and the mean service time. Hold speed and occupancy fixed, halve the variance term that measures irregularity, and the expected wait roughly halves. Halving the spread itself, in coefficient-ofvariation terms, would cut the wait to about a quarter.

The mining link is old too. Eliot Koenigsberg's 1958 work on cyclic queues applied queueing theory to underground coal mining, not to truck haulage. Cyclic and closed queue models have since become standard for truck-shovel systems. Engineers also size fleets with the 'match factor', which balances how fast a shovel fills trucks against how fast trucks complete a lap. What the desk adds is a direct claim: variability, not speed, is the lever autonomy pulls.

One honest limit. Kingman's formula was derived for an open queue fed from outside, not a fixed fleet circling a pit. So it guides intuition rather than settling the answer. A closed fleet needs simulation, which mine planners already run.

Some figures are public. Rio Tinto has reported about 12% higher productivity and about 15% lower load-and-haul unit cost (2017), as well as the extra 700 operating hours per truck per year. These appear in CIM Magazine, Mining Journal and emj.com. The 15% is a cost cut, not a productivity gain. We have not independently audited these company figures.

The most relevant outside evidence is a UBC simulation thesis comparing autonomous and manual fleets. Autonomous trucks were slower, and queuing time fell 28.7% with steadier driving. But the thesis credited the 21.3% production gain mainly to higher utilization, which is the hours explanation. That bears directly on our claim. It shows steadiness helps, but it does not show that variability is the main driver.

What we could not find is a published comparison of cycle-time spread, speed and tonnes per hour for human and autonomous fleets at one real mine.

The data already exist inside dispatch software, which timestamps every stage: queue at shovel, spotting, loading, travel, queue at crusher, dumping. Mines converted in stages, and Rio Tinto's autonomous trucks moved about a quarter of its Pilbara material in 2018 while the fleet expanded, which implies mixed human and autonomous operation. Anyone with access could compare, for each fleet, the spread of cycle times, average speed and tonnes per operating hour.

The desk states a scorable version: take the first qualifying published comparison of autonomous and human-driven haul trucks by June 30, 2028, reporting cycle-time spread, tonnes per operating hour and average speed for both at one mine. It passes if the autonomous trucks show lower spread, more tonnes per hour, and speed no higher than the human trucks. Conditional on such a comparison appearing, the desk gives the full claim a 60% chance. That figure is our own judgment, not a finding.

• The existing simulation. The one study we found attributes the production gain mainly to utilization, not to steadiness. If that holds in real mines, our thesis is at best a minor part of the story. • Confounding. Autonomous fleets arrive with new dispatch software, dedicated roads and fresh attention, any of which could lower variability on its own. • Spotting time. Robots may be slower to line up under the shovel, which would lengthen service and push the other way. This is plausible, but we found no source for it. • Access. The data is proprietary, and a firm may never publish a comparison showing its costly trucks are slower.

The health angle is indirect and is an analogy, not a finding. The same AI buildout that automates mines also automates hospitals and labs. When a system improves, the instinct is to credit speed or hours. Variability may be part of the real driver, and it is worth asking about: fewer erratic steps, more predictable throughput. That is one question to put to an AI tool in medicine, and to a claim that something 'got faster'.

• A published comparison reporting cycle-time spread, tonnes per hour and speed for both fleets at one mine. • Simulation studies of closed cyclic queues that isolate variability from speed and from hours. • Whether a 'steady beats fast' pattern shows up in automated clinical and lab workflows.

Do Driverless Mine Trucks Win by Being Steady, Not Fast? AI Agents Breached Hugging Face: What It Means for Medical AI Appeals Court Upholds Pentagon's Anthropic Label; Dispute Is About AI Safeguards