Safety Lessons from Billions of Autonomous Miles: What AV Deployment Teaches About Mining Autonomy
- May 12
- 6 min read
In the first two articles of this series, I covered how motion planning transfers across autonomous domains and why mining's constraints favor a modular architecture with learned intelligence and physics-based safety verification. But architecture is only half the story. The other half is proving that what you built actually works safely. The on-road AV industry has spent over a decade developing frameworks for exactly this problem, and mining has a lot to gain from studying what emerged.
What Two Hundred Million Miles Teaches You
Waymo recently crossed 200 million driverless miles on public roads. At that scale, the safety data starts to mean something statistically. The company's latest safety report shows 82% fewer injury-causing crashes and 13 times fewer serious-injury-or-worse incidents compared to human drivers across their operating cities. Swiss Re, one of the world's largest reinsurers, independently validated 92% fewer bodily injury claims over 25 million miles.
These numbers didn't come from building better sensors or faster processors. They came from building better safety practices around the technology. The frameworks for defining where a system can operate, how it gets validated, and what happens when it fails are what turned a research prototype into a commercial service.
The contrast with earlier AV efforts is instructive. In 2018, an Uber test vehicle struck and killed a pedestrian in Tempe, Arizona. The system had detected the person 5.6 seconds before impact but failed to classify a jaywalking pedestrian correctly, and emergency braking had been disabled in autonomous mode. Seven years later, a Waymo vehicle in Austin encountered a scooter rider who fell into its lane. The system had already begun a proactive avoidance maneuver before the fall was complete, decelerating and swerving with no human intervention. The difference between those two outcomes isn't just better algorithms. It is a more mature approach to defining, testing, and enforcing operational safety boundaries.
Three Frameworks Worth Borrowing
The AV industry has converged on a set of safety practices that aren't specific to passenger cars. They address a general problem: how do you certify that an autonomous system will behave safely in a defined environment? Mining faces this exact problem.
Operational Design Domains
An ODD defines the specific conditions under which an autonomous system is designed to operate: geographic boundaries, weather conditions, road types, speed ranges, time of day, and the types of objects it must handle. The concept seems obvious, but its power is in forcing explicit specification. Instead of claiming "this system drives safely," you claim "this system operates safely on surface streets in Phoenix, in dry conditions, at speeds below 45 mph, during daylight and nighttime." Every condition is documented. Every boundary is testable.
For mining, ODD thinking translates directly. A haul truck's ODD might specify: surface operations on graded haul roads, grades below 12%, in the presence of other autonomous trucks and designated light vehicles, with GPS availability above a defined threshold, and wind speeds below a specified limit. Making these boundaries explicit does two things. It gives mine operators a precise understanding of where autonomy applies and where human operation is still required. And it creates a testable specification against which the system can be validated.
Safety Cases (UL 4600)
The AV industry's UL 4600 standard takes a different approach from traditional prescriptive safety standards. Rather than listing specific technical requirements, it requires the developer to build a structured safety argument: a set of claims about what the system does safely, supported by evidence, organized into a traceable chain of reasoning. The standard is goal-based and technology-agnostic, meaning it applies equally to LiDAR-based systems and camera-only approaches, to modular and end-to-end architectures. Its scope explicitly covers autonomous trucks, tractors, construction equipment, and mining vehicles. Edition 3 expanded coverage for autonomous trucking, and the framework applies directly to off-road industrial machinery.
Mining already has ISO 17757, which covers safety requirements for autonomous and semi-autonomous machines in earth-moving and mining. But ISO 17757 addresses hazard categories and safety principles at a high level. It doesn't prescribe how to build the detailed, evidence-backed safety argument that operators and regulators increasingly expect. Adopting UL 4600's safety case methodology alongside ISO 17757 would give mining operators a structured way to demonstrate safety that goes beyond compliance checklists.
Transparent Safety Reporting
NHTSA's Standing General Order, first issued in 2021, requires AV manufacturers to report crashes involving autonomous systems within days. This mandatory reporting has created a public dataset that researchers, insurers, and regulators can independently analyze. Waymo has gone further, publishing peer-reviewed safety studies and maintaining an open Safety Impact Hub with its crash data.
Mining generally keeps safety data internal. Incident reports go to mine safety regulators, but aggregated safety performance data for autonomous haulage is rarely published in a way that allows cross-operator comparison. If the mining industry adopted something like the AV sector's transparent reporting model, even voluntarily, it would accelerate the industry's ability to benchmark autonomous safety and identify systemic issues across sites.
Why Mining Has an Advantage
Here is what makes mining's safety challenge different from, and in some ways easier than, the on-road problem.
On public roads, the ODD is vast and the edge cases are nearly infinite. A robotaxi must handle construction zones, emergency vehicles, pedestrians on phones, double-parked delivery trucks, and an ever-changing built environment. Validating safety across that space requires hundreds of millions of miles and massive simulation investments.
A mine site is a controlled environment. The roads are private. The traffic is managed. The agents are known. The ODD is bounded in ways that public roads never will be. This means the validation problem, while still hard, is more tractable. You can enumerate scenarios more completely, test against the full ODD more realistically, and enforce ODD boundaries through physical infrastructure and operational procedures in ways that public roads don't allow.
But a controlled environment also introduces a risk that public roads don't have: complacency. The very fact that mining operations are structured and predictable can create overconfidence in safety outcomes. Formal safety frameworks guard against this. An ODD specification forces you to document every assumption. A structured safety case requires you to back every claim with evidence. Transparent reporting creates accountability that internal reviews alone can't provide. These frameworks are valuable not because mining's safety problem is harder than the on-road problem, but because they impose discipline on a problem that might otherwise feel easier than it actually is.
This is an advantage mining should press. The frameworks developed for the hardest version of the autonomy safety problem are being offered to an industry where the safety problem is more constrained and more controllable. The question isn't whether mining can afford to adopt these frameworks. It is whether mining can afford not to.
Some of this adaptation work is already underway. The Association of Equipment Manufacturers (AEM), which represents the off-road equipment industry across agriculture, construction, and mining, has been engaging with California's Cal/OSHA advisory committee on safety regulations for autonomous agricultural equipment. A key tension in that process has been pushback against importing the California DMV's on-road autonomous vehicle framework directly into off-road applications. AEM has also published an autonomy standards development whitepaper and coordinates industry participation in standards like ISO 18497 (safety of autonomous agricultural machinery) and ISO 3502 (advanced automation in mining systems). This is the right kind of work: taking the safety thinking from the AV industry seriously while insisting that the implementation reflects the realities of off-road operations.
A Safety Culture, Not Just Safety Features
The deepest lesson from the AV industry isn't any single framework or standard. It is that safety in autonomous systems is a culture, not a feature. It is built into how you define what the system should do, how you test whether it does it, how you monitor performance after deployment, and how you respond when something goes wrong.
Mining has a strong existing safety culture. The zero-harm philosophy and the rigor of mine safety regulation provide a foundation that many other industries lack. The opportunity is to bring the AV industry's technical safety methodology into that existing culture: ODD specification, structured safety cases, scenario-based validation, and transparent performance reporting.
The autonomous vehicle industry learned these lessons over billions of miles and more than a decade of deployment. Mining doesn't need to repeat that journey from scratch. The frameworks exist. The data is public. The transfer is waiting to happen.
