The Seeing Mine: How Physical AI Is Changing Real-Time Decision-Making in Mining

In discussions about AI in mining, attention usually goes to analytics: predictive maintenance models, production forecasts, dashboards that explain the last shift. It’s valuable work, but it all happens after the fact. The harder problem sits on the haul road, at the face and around the crusher, where conditions change by the minute and a late response is measured in injuries and lost tonnes.

Physical AI is aimed at fixing this problem. The term ‘Physical AI’ covers systems that read the physical environment through cameras and sensors, process what they see on site, and support a decision within seconds rather than at the next shift meeting. In an industry where people, heavy equipment and moving material share the same space, that capability has been a long time coming.

Ramjack has partnered with alwaysAI, a specialist in edge computer vision, to bring Physical AI into live mining operations.

The Real-Time Mandate

Most mines already have cameras, often hundreds of them. Very little of that footage is watched live, and most of it is only reviewed after something has gone wrong. A vision model running on a small device on the network can change everything. It can recognise a person inside an exclusion zone, a vehicle in the wrong lane, a belt tracking off-centre or spillage building at a chute and raise an alert while there is still time to act among many other things.

What this really changes is the speed of decision-making. On most sites, the time between something happening and someone acting on it depends on how quickly the information reaches the right person. A supervisor reads an incident report the next morning. A control room operator scanning dozens of feeds may notice a problem minutes later, or not at all. Physical AI shortens that gap. It picks out what matters from what the cameras see, adds the location and context, and puts it in front of the person or system that can act, whether that means stopping a conveyor, redirecting a haul truck or clearing an area.

It also improves decisions over time. Every event is logged with a time, a location and footage, so supervisors and engineers can see patterns, such as repeated near-misses at the same intersection or spillage that builds after a particular shift change, and deal with the underlying cause.

Processing on site is important because mine networks have limited bandwidth and coverage is rarely complete. A system that sends video to the cloud for analysis will be at its slowest when speed counts most. With edge processing, only events and short clips need to travel across the network.

The stakes are high on both the safety and production side. Mobile equipment and transport were the leading cause of fatalities among ICMM member companies in 2024, accounting for nine of the 42 deaths reported. Meanwhile, operators at one of the world’s largest mines estimate that a single conveyor belt failure can cost between US$6 million and US$12 million. In many of these incidents, the warning signs are visible well before the event, provided someone, or something, is watching for them.

The Ramjack and alwaysAI Partnership

alwaysAI provides a system for building, deploying and managing computer vision based physical AI applications on edge hardware. Ramjack brings more than a decade of operational technology integration in mining, from site networks and monitoring systems to control rooms and remote operations through rROC. Together, the two companies make sure that what a camera detects reaches the people and systems that need to respond.

Why Mines Struggle to Get Physical AI Right

If the case for Physical AI is so clear, why are so few mines using it at scale?

The first reason is fragmentation. Mines already run collision avoidance, access control, fleet management, environmental monitoring and plant control, usually from different vendors. A vision system bolted on as one more standalone screen produces alerts that nobody owns. If a model detects a person near a moving conveyor but has no link to the conveyor’s control system or the control room, it has only done half the job.

The second reason is the environment itself. Light changes through the day, dust levels shift with weather and blasting, and pit and underground layouts change constantly. A model that performs well in a demonstration still needs tuning to real site conditions, and ongoing attention as the operation evolves.

The third reason is infrastructure. The AI needs to be optimised to the Edge and needs to constantly adapt, in real-time to the dynamic network constraints. Edge processing reduces the load on the network but does not remove it. Alerts and footage still need a reliable route to the people who act on them, and many sites have coverage gaps in exactly the places cameras are most needed.

The Ramjack Approach

At Ramjack Technology Solutions, we treat computer vision as part of the operation from day one. As a vendor-independent OT integrator, our role is to make it work alongside the systems already on site.

• Site Assessment & Use-Case Selection: We start with the operation’s own priorities, whether safety, equipment availability or security, identify where vision will deliver measurable value, and make use of existing cameras wherever possible.
• Edge-to-System Integration: We connect alwaysAI’s vision applications to the platforms the mine already runs, including control rooms and rROC, fleet and collision avoidance systems, plant control and reporting, so that a detection leads to a decision and an action.
• Deployment & Ongoing Optimisation: We install, tune and maintain the models and the network that supports them, so performance holds as site conditions change.

Realising the Seeing Mine

Most Physical AI deployments today detect and alert, and leave the decision with a person. The next step will be automatic responses, within rules that operators set and trust, such as slowing a vehicle, stopping a belt or adjusting ventilation, with people focusing on the exceptions. Further ahead, vision will combine with other sensor data and with autonomous equipment, giving machines a shared, live view of the site they work in. At every stage the aim is the same: to shorten the time between something happening on site and the right decision being made about it.

The operations that benefit most will be the ones that treat computer vision as part of how the mine runs, integrated and maintained over time, rather than as another pilot.

If you are looking at computer vision for safety, performance or security, Ramjack and alwaysAI would welcome the conversation.

Real-time decisions start with seeing. Don’t run tomorrow’s mine on yesterday’s footage.