What is mining autonomy data collection?
Mining autonomy data collection is the capture of sensor data and operator demonstrations that train autonomous haul trucks, underground loaders, drill rigs, and inspection robots. It means LiDAR, camera, thermal, and IMU data recorded in dust, darkness, and GPS-denied tunnels, plus teleoperated demonstrations from skilled operators, packaged with the calibration and condition tags a mine-grade model needs.
Surface haulage has already proven autonomy works in mining. Caterpillar reported that in 2025 it expanded its autonomous haul truck fleet to 827 trucks in operation and signed its first mixed-fleet autonomy agreement.
What that means for the next wave: the easy open-pit routes are taken. The growth is underground, in loading and drilling, and in inspection, where sensors struggle most and public data barely exists.
Where mining autonomy needs better data
| Application | What the system must learn | Data we collect | Services |
|---|---|---|---|
| Surface haulage perception | Small hazards next to huge machines: people, light vehicles, rocks | Truck-mounted LiDAR and camera runs by shift, weather, and road condition | LiDAR, edge case |
| Underground loaders (LHD) | Localization in tunnels, bucket loading in muck piles | LiDAR, IMU, and camera runs after blasting, in dust and wet ground | LiDAR, edge case |
| Drilling and charging | Hole alignment, rod handling, charging hoses | Teleoperated demonstrations from skilled operators with force and vibration data | Teleop, multimodal |
| Stope and void inspection | Geometry of spaces people can't enter | Handheld, vehicle, and drone LiDAR scans with survey control | Point cloud, aerial |
| Conveyor and plant inspection | Idler failures, belt damage, hot bearings | Camera, thermal, and audio capture along conveyors, with failure examples | Multimodal, edge case |
The Harsh Envelope: the conditions a mine dataset can't skip
Gamasome framework
Mines break sensors in ways other industries don't. The Harsh Envelope lists the five conditions every mining capture plan has to cover on purpose, because they're the ones that fail autonomy in production.
In a mine, the clean frames are the edge cases.
Dust and particulates
Post-blast dust, haul road dust, and diesel haze, recorded with lens-state and visibility tags rather than scheduled around.
GPS-denied positioning
Underground runs captured with survey control and reference poses so localization can be trained and proven without satellites.
Vibration and shock
Sensor mounts on heavy equipment shift. We log extrinsic checks per session so drift is caught before it corrupts a batch.
Scale mismatch
A 300-ton truck must see a person or a pickup. Collection plans compute points on target for the smallest hazard, not the biggest machine.
Low light and glare
Headlights, cap lamps, and total darkness underground, plus low sun on open-pit benches.
What an underground data program looks like
Composite scenario: autonomous underground loader (details generalized)
- Situation
- An underground loader runs autonomously on its main haul route but hands back to a remote operator near the draw points.
- Problem
- Draw points are where dust peaks after blasting and muck piles change shape every cycle. The training data came mostly from clean tunnel runs between them.
- Solution
- Record LiDAR, IMU, and camera data at draw points across blast cycles, pair it with teleoperated bucket loads from experienced operators, and tag dust level and pile state on every run.
- Outcome
- A dataset centered on the zone where autonomy was dropping out, plus held-out draw points from a second level to test before rollout.
For the mine's operations manager, the metric is simple: fewer handbacks to the remote operator per shift. Data targeted at the handback zone is how that number moves.
Common mistakes in mining autonomy data
Recording between blasts only
The worst dust follows blasting, which is exactly when autonomy struggles. Capture through the cycle.
Unchecked sensor mounts
Vibration moves sensors a little every shift. Without extrinsic checks, a month of data can drift silently.
Big-object bias
Datasets full of trucks and loaders underrepresent people and light vehicles, the hazards that matter most.
Ignoring operator expertise
Skilled operators load buckets and position rigs in ways that are hard to script. Their teleoperated demos are worth capturing.
All mining work runs under the site's safety management system and inductions. See how LiDAR programs handle dust and range in our LiDAR data collection service.
Mining autonomy data FAQs
What is mining autonomy data collection?
It's the capture of sensor data and operator demonstrations that train autonomous haul trucks, underground loaders, drill rigs, and inspection robots, recorded in real mine conditions such as dust, darkness, and GPS-denied tunnels.
How do you collect data underground without GPS?
We capture LiDAR, IMU, and camera data alongside survey control points and reference poses, so localization models can be trained and verified without satellite positioning.
How do you handle dust in mining data?
We treat dust as a condition to cover, recording across blast cycles and haul road conditions and tagging visibility and lens state on every run.
Can you capture demonstrations from skilled mine operators?
Yes. Teleoperated demonstrations from experienced operators for loading, drilling, and positioning are recorded with the machine's own states and sensor streams.
Do you work under mine safety rules?
Yes. All collection runs under the site's safety management system, inductions, and access procedures, scheduled with operations.
Can you scan areas people can't enter?
Yes. Stopes, voids, and hazardous areas can be scanned with vehicle-mounted, handheld, or drone LiDAR, registered to survey control.