Mines and quarries move enormous tonnages with large fleets of heavy machines on private, geofenced haul roads. Autonomous haulage systems (AHS) replace the driver with on-board perception and control, coordinated by a central fleet manager and watched from a remote operations centre. Operators report steadier cycle times, better utilisation and fewer people in the line of fire; the cost is a high-integration, safety-critical programme rather than a software purchase.
Problem & value
- Today: haul trucks, loaders and drills are driven by operators in 12-hour shifts, in dust, night and bad weather. Hauling is often the largest single operating cost of a surface mine. Fatigue, shift changes and variable driving style hurt productivity and safety.
- AI value: consistent driving, fewer shift-change losses, fleet-wide dispatch optimisation, less wear (smoother driving, fuel/energy use), and removal of people from the most dangerous zones. Remote operators can supervise many machines from one site or from another country.
- Scale evidence: Komatsu announced in April 2026 that it had commissioned its 1,000th autonomous ultra-class haul truck and reports FrontRunner customers having hauled over 11.5 billion tonnes (company-reported) [1]. Caterpillar (MineStar Command for hauling) and others report comparable large fleets in their own publications; figures are vendor-reported and not independently audited.
- Europe is smaller but real: autonomy in EU sites is concentrated in underground and quarry niches (autonomous drilling and loading in Swedish and Finnish mines, autonomous quarry haulers from Volvo CE and others), while the largest open-pit fleets are in Australia, Chile and North America.
Variants
- Open-pit haulage: ultra-class trucks (Komatsu FrontRunner, Caterpillar MineStar Command, Hitachi/Wenco-based systems, Liebherr and others) on long, well-maintained haul roads.
- Quarry and aggregates haulage: smaller articulated or rigid trucks on short, steep cycles, often retrofitted (e.g. Volvo CE, Sandvik/Epiroc-class equipment, third-party kits).
- Autonomous drilling: blast-hole drill rigs positioning and drilling from a drill plan under remote supervision (Epiroc, Sandvik, Cat, Komatsu/Joy).
- Underground loading and haulage: LHDs and trucks with tunnel-referenced navigation (Epiroc, Sandvik) where there is no GNSS.
- Loaders, dozers and yard tractors: semi-autonomous loading, stockpile and waste-dump work; autonomous terminal tractors and trucks on closed port, steel or logistics yards (see ROBO-014, ROBO-020 for adjacent cases).
- Retrofit autonomy kits: third-party stacks (e.g. Pronto, SafeAI, Teleo-style teleoperation and autonomy kits) for mixed fleets.
How it works
- machine sensors (lidar, radar, camera, GNSS/INS, wheel/IMU)
- localisation on site map + perception (obstacles, road edge, berms, dust-robust)
- on-board planning & control (path following, speed, load/dump manoeuvres)
- safety layer (independent obstacle/zone monitoring, e-stop, ISO 17757 logic) ⇅ wireless network (private LTE/5G or mesh) ⇅ fleet manager: dispatch + traffic management + geofence/exclusion zones
- coupled with mine planning (shovel/crusher queues, grade, blend)
- remote operations centre: supervision, assist, recovery, exceptions
- Perception: lidar/radar-centred object detection and terrain segmentation with learned components; dust, rain and glare degrade cameras and lidar, so radar and map-based localisation carry much of the load. Learned models detect unexpected objects, berm edges and manned vehicles; deterministic rules keep the safety function predictable.
- Fleet level: dispatch is an optimisation problem (assignment of trucks to loaders, routing, queueing) solved with operations research plus ML estimates of cycle times. Autonomy is managed per haul-road segment; manned and unmanned vehicles are separated by zones, traffic rules and proximity detection.
- Human-in-the-loop: remote operators handle exceptions, unusual obstacles, loading in tight spots and recovery. The safety-rated stop function and the decision to release an area to autonomous operation stay with defined human roles and procedures.
Data preparation
Data comes from the machines (lidar, radar, camera, GNSS/INS, telemetry), the mine planning and fleet management systems, surveyed site maps and geofences, and personnel or vehicle tag systems. In a typical operator it is spread over OEM telematics portals, historians (SCADA), survey files and spreadsheets, and rarely in ERP or a DMS. Typical gaps: outdated or inconsistent haul-road maps, no unified machine and event IDs across vendors, unlabelled near-miss events, uneven network and positioning coverage, and tag data that identifies individual people.
Preparation steps: - Survey and version the site map, geofences and speed zones; keep one authoritative copy. - Export machine, dispatch and health data via APIs or historians (PREP-005) and align IDs and time stamps (PREP-011, PREP-012). - Build evaluation scenarios from recorded safety events with domain experts (PREP-018, PREP-020). - Pseudonymise personnel tag data used for analysis (PREP-022). - Machine health data can feed PRED-006 in parallel.
With Path A the vendor handles perception training, but the operator must deliver clean maps, interface data and event labels; plan roughly 30-40 percent of the integration effort for it. For Path C this rises to around half, plus labelling of site-specific scenes.
Implementation paths
Which path for whom
- Enterprise (mining, quarry, port and yard operators): Path A, because mature OEM and third-party autonomous haulage systems exist and the safety case, regulator consultation and service are carried by the vendor.
- Developer (robotics and autonomy vendors, research): Path C, because open perception, planning and simulation stacks allow building or extending autonomy for specific machines or closed sites.
Path A — Buy (default)
- Turnkey OEM systems: Komatsu FrontRunner, Caterpillar MineStar Command, Hitachi Construction Machinery with Wenco, Liebherr Autonomous Haulage, Volvo CE autonomous solution (EU), Epiroc and Sandvik autonomy for drilling and underground (EU). Each bundles trucks or retrofit, fleet manager, communication and safety case.
- Retrofit and mixed-fleet vendors: e.g. Pronto, SafeAI, Teleo, ASI Mining; fleet management and dispatch from Modular Mining (Komatsu), Wenco, Epiroc, Sandvik.
- Choose A when: you operate a mine or quarry with enough haul volume and long-lived routes; for most SMB quarries start with assistance functions (collision avoidance, loader assist) rather than full autonomy.
Path B — Build on cloud APIs
- Not applicable for the vehicle control loop: latency, availability and safety rules keep it on-board and on-site. Cloud APIs fit around it: digital-twin analytics, fleet KPIs, predictive maintenance (PRED-006), map change detection, and LLM assistants for shift reports and incident summaries (EU-region hosting, e.g. Azure, AWS Frankfurt, Mistral).
Path C — Self-host (open source)
- Research and pilot stacks: ROS 2 with Nav2, Autoware (originally for roads, adapted to off-road), open lidar-SLAM packages, OpenVDB/Open3D for terrain, and open perception models (YOLO-class and DETR-class detectors, point-cloud networks).
- Simulation: Gazebo, NVIDIA Isaac Sim, CARLA-style tooling for scenario tests (see ROBO-010).
- Hardware: on-vehicle industrial edge computers or automotive-grade GPUs (HW1 class, e.g. Jetson Orin); control-room servers for fleet and dispatch.
- Reality check: open stacks have no safety certification for 100-t machines; they suit research, small vehicles and teleoperation-assist, not production heavy haulage.
Path D — Custom training
Worthwhile for site-specific perception (local rock, dust and lighting, unusual obstacles), learned cycle-time and dispatch models, and sim-to-real policies. Needs months of recorded multi-sensor data from your own machines, labelled edge cases (people, light vehicles, animals, fallen rock), and a scenario library for regression testing.
State of the art (2026)
- Mature at scale for haul trucks on defined roads: OEM systems run fleets of hundreds of trucks; the first generation was commercial in 2008 [1]. Productivity gains are vendor-reported and depend heavily on site layout and baseline.
- Frontier: mixed fleets of OEMs under one fleet manager, interoperability standards (e.g. the Mining Interoperability work around ISO 17757 and AEMP/ISO 15143), autonomous drilling and loading, battery-electric and trolley-assist haulers with autonomy, and teleoperation with learned assistance.
- Learned perception supplements rule-based safety; large vision-language models are used offline (data triage, incident review), not in the safety loop.
- Limits and failure modes: dust, fog and snow; changing terrain (berms, ramps, wet haul roads); GNSS multipath in deep pits; wireless network outages; unusual obstacles; interaction with manned light vehicles; remote-operator handover delays; cybersecurity of the fleet network.
Shared resources & combinations
- Perception and edge stack shared with ROBO-005 (road AD), ROBO-007 (off-road agriculture) and ROBO-015 (plant inspection robots).
- Safety functions (person detection, zone monitoring): ROBO-016. Worker proximity and PPE: VIS-007.
- Simulation and sim-to-real: ROBO-010. Terrain and ground-motion monitoring: VIS-061.
- Maintenance: PRED-006 predictive maintenance on telemetry; drones for stockpile surveys: ROBO-006.
- Fleet optimisation (optimization-solver) is shared with ROBO-002 and logistics use cases.
Effort, cost & KPIs
- Path A: 12-36 months from business case to a scaled fleet: site survey, network build-out (private LTE/5G), safety case, regulator consultation, change management for operators. Cost drivers: trucks or retrofit kits, network and positioning, control room, integration with mine planning, training, ongoing OEM service.
- Path C: research or pilot scale only; staff cost dominates.
- KPIs: autonomous hours and availability, tonnes per hour, cycle time and queueing, fuel/energy and tyre cost per tonne, interventions per 1,000 hours, safety events and near misses, remote-operator-to-machine ratio, network uptime.
People & change
Start with management and the operations or mine manager (business case, safety targets), then bring in IT and IT security (private LTE/5G, remote access, OT segmentation), legal/compliance (mining authority, machinery safety, liability) and the works council before pilot contracts are signed. Drivers and dispatchers are the end users most affected: involve them as remote operators, supervisors or maintenance staff early.
Works council: yes. Telemetry, tag positions and remote-operator logs can show individual behaviour and performance, so §87(1) no. 6 BetrVG applies. Reorganised work processes with AI trigger information duties under §90(1) no. 3; consider an AI expert under §80(3). Public operators involve the Personalrat. Seek a works agreement on purpose, no individual performance scoring and access rights.
Personal data: tag-based location of staff requires the DPO and likely a DPIA (Art. 35 GDPR) plus an AVV (Art. 28 GDPR) with vendors.
AI literacy: support the development of operators' AI literacy (Art. 4 AI Act) with training on system limits, intervention and emergency stop. Communicate job impacts honestly, offer retraining paths, start with one pit or zone and a volunteer pilot group, and keep humans in control through remote supervision and independent safety functions (ROBO-016). Guides: GUIDE-004, GUIDE-008, GUIDE-009.
Risks, regulation & compliance
- EU AI Act: AI that is a safety component of machinery falls under Annex I (machinery legislation, now Regulation (EU) 2023/1230, applying from 20 January 2027) and is therefore high-risk under Art. 6(1), with obligations applying from 2 August 2028 (postponed from August 2027 by the Digital Omnibus on AI, Reg. (EU) 2026/1744) [2][3]. Not an Art. 5 prohibited practice, but worker monitoring (driver-behaviour scoring, emotion recognition at work) is a separate question: emotion recognition at work is prohibited under Art. 5.
- Machinery and safety: Regulation (EU) 2023/1230 (conformity assessment, safety-related control systems, cybersecurity), ISO 17757 for autonomous machine systems in earth-moving and mining, ISO 13849 and IEC 62061 for safety functions, IEC 61508 as basis. Mining authority rules (national, e.g. BG RCI and Bergbehörden in Germany) govern site approval and personnel exclusion [2].
- GDPR: mostly machine data, but camera feeds, operator voice/video in the control room, worker tags and location data are personal data; works council co-determination (e.g. Germany BetrVG §87) applies to monitoring. Define purposes, retention and access [4].
- Cybersecurity: the fleet network is critical infrastructure on site; apply IEC 62443 and the Cyber Resilience Act; NIS2 may apply to large operators.
- Mitigations: independent safety layer, exclusion zones, tested recovery procedures, staged rollout (shadow mode, then supervised, then driverless), incident review with logged AI decisions.
Getting started
- Baseline haul cost per tonne, cycle times, incidents and operator availability; pick a candidate haul route with stable layout.
- Run a feasibility study with 2-3 vendors (OEM and retrofit) against your fleet and site; check the regulator's requirements early.
- Build the foundations: private wireless network, positioning (GNSS plus reference), fleet management data, geofences.
- Pilot on one route or bench with a small fleet, in shadow then supervised mode; define remote-operator roles and shift model.
- Involve workers' representatives and set up data-protection documentation before launch.
Sources
- Komatsu — Komatsu becomes first OEM to commission 1,000 ultra-class autonomous haul trucks (Apr 2026) — https://www.komatsu.com/en-us/newsroom/2026/komatsu-becomes-first-oem-to-commission-1000-ultra-class-autonomous-haul-trucks
- Regulation (EU) 2023/1230 on machinery — https://eur-lex.europa.eu/eli/reg/2023/1230/oj
- Regulation (EU) 2024/1689 (AI Act) — https://eur-lex.europa.eu/eli/reg/2024/1689/oj
- Regulation (EU) 2016/679 (GDPR) — https://eur-lex.europa.eu/eli/reg/2016/679/oj
- ROS 2 Navigation (Nav2) — https://github.com/ros-navigation/navigation2
- Autoware Foundation — https://github.com/autowarefoundation/autoware
- Komatsu newsroom, April 2026 — https://www.komatsu.jp/en/newsroom/2026/20260422