A mine is about as far from a data centre as an AI system gets. Pits and underground workings are often hours from a city, sit in heat, dust and vibration, and can't pause because a network link dropped. That's why most AI that matters on a mine site runs on the machine, the camera or a rugged box nearby, not in the cloud.
Why mining pushes AI to the edge
- Coverage can't be assumed. The 2024 Regional Telecommunications Review reported recent progress but also ongoing connectivity challenges in regional and remote Australia. Underground, satellite positioning disappears entirely, which is why systems such as Emesent's Hovermap are built for GPS-denied environments.
- Safety decisions are local. A warning about a person behind a reversing machine is useless if it arrives after a round trip to a server.
- Video is heavy. Several high-resolution cameras per machine, across a fleet, generate far more data than most site links can carry. Sending detections instead of footage is the practical answer.
- The environment is hard on electronics. Robotic Systems, for example, says it builds bespoke hardware optimised for the power, temperature and vibration requirements of mining.
Where edge AI is used on Australian mine sites
Vehicle interaction and blind spots
Interactions between vehicles, mobile plant and people are among the most serious hazards in mining. The NSW Resources Regulator treats roads and other vehicle operating areas as a principal hazard, and in July 2025 released a technical reference guide to help surface mines reduce adverse vehicle interactions. Camera-based detection is one tool operators use alongside other controls. Presien, a Sydney company that works with heavy industry, integrates AI directly into machines. Its Blindsight system gives operators in-cab alerts about their blind spots, with video-based reporting and a risk score for safety teams.
Mapping and survey where GPS can't reach
Underground drives, stopes and voids are dangerous to survey by hand. Emesent, a Brisbane spin-out from CSIRO, says its Hovermap enables autonomous underground and open-pit mapping in GPS-denied conditions, and lists stope reconciliation among its solutions. The autonomy has to run on the platform itself, because a drone inside a stope can't wait for instructions from the surface.
Fixed plant: conveyors, crushers and processing
Cameras and sensors on fixed plant can flag spillage, blockages, belt damage, foreign objects or abnormal vibration. These jobs suit edge AI because they run around the clock and produce huge volumes of mostly uneventful data. Robotic Systems develops machine learning for vision, signal and spatial data that works on the edge with no cloud dependency.
Custom sensors, robots and inspection
When no product fits, the work becomes an engineering project: a new sensor, a custom board, a model trained on site data, or a robot for a task people shouldn't be doing. Vedametric, which works across Sydney, the Gold Coast and Brisbane, designs boards around edge AI accelerators, compresses models to run offline in the field, and trains robots in simulation before transferring them to real hardware.
Engineering constraints that catch projects out
- Dust, water and cleaning. Lenses get coated and hosed down. Plan to detect a dirty or blocked camera, not just the target objects.
- Heat. Sealed enclosures in direct sun throttle processors. Test in the hottest conditions the site sees; the hardware guide covers power and thermal trade-offs.
- Vibration and shock. Connectors, cables and mounts usually fail before the silicon does.
- Power. Machine-mounted systems run from the vehicle supply and must cope with starting, stopping and isolation.
- Change control. Anything that alerts operators or touches machine control will go through the site's change management and risk assessment. Allow time for it.
- Maintenance access. If a fix needs someone to drive to a remote pit, design for remote diagnostics and updates from the start.
Data is usually the bottleneck
A model for a mine site needs images and signals from that site, or one very like it. Public datasets rarely show haul roads at dusk, dust-covered high-vis or diesel haze. And the events that matter most, near misses and equipment failures, are rare by definition. Practical steps:
- Start recording representative data early, before the model exists.
- Capture across shifts, seasons and weather, and label conditions as well as objects.
- Run new systems in shadow mode first: the model makes and logs predictions but nobody is alerted, so false alarms can be measured before operators see any.
- Decide what footage of workers you keep, why and for how long. The on-device AI and privacy guide covers the Australian Privacy Principles.
Australian firms with mining-relevant edge AI
| Firm | Base | What it offers mines |
|---|---|---|
| Robotic Systems | Newcastle, NSW | Rugged custom sensing, machine-vision and control products; advertises business case to working prototype in 4 months; states 100% customer-owned IP |
| Emesent | Brisbane, QLD | Hovermap autonomous LiDAR mapping and the Cortex autonomy engine for GPS-denied spaces |
| Presien | Sydney, NSW | Blindsight on-machine blind-spot detection with in-cab alerts and safety reporting |
| Vedametric | Sydney, NSW | Custom edge AI: model training and compression, custom boards, offline inference and robotics |
Much of this capability traces back to public research. Emesent began as a spin-out from CSIRO's robotics group in Brisbane, whose field robotics work spans mining, agriculture and manufacturing.
Questions to ask a supplier for a mine-site system
- Has the hardware been tested for the temperature, dust, water and vibration on this site, and what is its rating?
- Can the system run with no network at all, and what does it store while it's offline?
- How are model and firmware updates delivered to machines in remote pits or underground?
- What happens when a camera is dirty, blocked or misaligned? Does the system report it?
- Which evaluation data was used, and does it include footage from conditions like ours?
- Who owns the model, the training data and the hardware design at the end of the project?
How a mining edge AI project usually runs
- Feasibility. Define the hazard or inefficiency, the decision the system supports and how success will be measured. Collect sample data.
- Proof of concept. Train a first model and test it on recorded site data.
- Shadow trial. Install on one or two machines, or one section of plant, logging only.
- Supervised operation. Switch on alerts, with close review of every event.
- Rollout. Production hardware, fleet updates, and training for operators and maintainers.
Robotic Systems advertises going from business case to working prototype in four months. The full path to a fleet rollout depends on data, hardware and site approvals.
FAQ
- Can edge AI work underground without network coverage?
- Yes. Inference runs on the device, so it doesn't need a link to make decisions, and results and logs can be stored and uploaded when the machine is back in coverage. Positioning is the harder part: underground systems such as Emesent's Hovermap are designed for GPS-denied conditions.
- Does camera-based AI replace proximity detection or collision avoidance systems?
- Treat it as one control among several, assessed within the site's own principal hazard management plan rather than as a replacement. Check current guidance from your state regulator, such as the NSW Resources Regulator's material on roads and other vehicle operating areas, and involve qualified safety professionals.
- How long does a mining edge AI pilot take?
- It depends on data, hardware and site approvals. Robotic Systems advertises business case to working prototype in four months. Shadow trials and site change management add time before alerts reach operators.
- Who owns the IP in a custom mining AI system?
- Your contract decides. Some firms state their position up front; Robotic Systems, for example, says clients keep 100% of the IP. Settle ownership of the model, training data, board design and firmware before work starts.