We’re putting $5.5M behind research and development of smaller, specialized AI models for real-world deployment — Read our manifesto →
    All posts

    [Case Study] Autonomous Haulage at Mine Scale: The Shougang and EACON Case

    How mixed-fleet Level 4 autonomy combined perception, vehicle communication, and dispatch to lower reported cost per tonne and reduce high-risk human work.

    Conscious Engines

    Industry: Mining
    Organizations: Shougang Group and EACON Mining
    Use case: Level 4 autonomous haulage and integrated dispatch
    Evidence basis: World Economic Forum MINDS cohort reporting, with Komatsu scale context
    Disclosure: This is an independent analysis by Conscious Engines. The principal outcome figures are participant supplied through a World Economic Forum program.

    1. Outcome at a Glance

    The Shougang and EACON deployment supports more than 70 truck models in a mixed-fleet autonomous haulage environment. The public account reports a 90% reduction in cost per tonne, a 13.3% reduction in energy use, and an 82.5% reduction in high-risk human workload.

    Key Outcomes

    More than 70

    Truck models supported

    Mixed-fleet compatibility reported in the cited case.

    90% reduction

    Cost per tonne

    Participant-reported result.

    13.3% reduction

    Energy use

    Participant-reported result.

    82.5% reduction

    High-risk human workload

    Participant-reported result.

    MeasureReported resultEvidence note
    Truck models supportedMore than 70Mixed-fleet compatibility
    Automation levelLevel 4Operational design domain applies
    Cost per tonne90% reductionParticipant-reported result
    Energy use13.3% reductionParticipant-reported result
    High-risk human workload82.5% reductionParticipant-reported result

    The cost reduction is unusually large and requires diligence. The public article does not provide the full baseline, labor and capital treatment, site boundaries, or calculation period. It should be treated as a reported case result, not a universal autonomous-haulage benchmark.

    For market context, Komatsu announced commissioning its 1,000th ultra-class autonomous haul truck in 2026 and reported more than 11.5 billion tonnes hauled by its autonomous fleet. That separate evidence shows autonomy at global mining scale, although it does not validate the Shougang economics.

    2. The Operational Problem

    Mine haulage combines large mobile equipment, repeated routes, variable road and weather conditions, and significant safety exposure. Human operators perform long shifts in high-risk environments. Inconsistent speed, braking, queuing, and dispatch can increase energy and reduce production.

    Autonomy must work as a fleet system. A truck cannot optimize itself independently if shovels, crushers, roads, fuel points, and other vehicles are constrained. Mixed fleets add different braking, payload, sensor, and control characteristics.

    The operating risk is physical. A model error can damage equipment or harm people. The design must therefore use redundant perception, constrained behavior, geofencing, emergency systems, deterministic safety controls, and a clearly defined operational design domain.

    3. What Was Built

    The public description includes high-precision positioning, 360-degree perception, vehicle-to-vehicle communication, and integrated dispatch.

    System at a Glance

    Positioning

    Locate vehicles precisely on mine roads.

    Perception

    Detect road, obstacles, people, and equipment around the truck.

    Vehicle communication

    Share intent and state across the fleet.

    Motion control

    Execute safe speed, steering, braking, and stopping.

    Fleet dispatch

    Coordinate assignments, queues, and production targets.

    Control room

    Monitor exceptions and enable intervention.

    LayerFunction
    PositioningLocate vehicles precisely on mine roads
    PerceptionDetect road, obstacles, people, and equipment around the truck
    Vehicle communicationShare intent and state across the fleet
    Motion controlExecute safe speed, steering, braking, and stopping
    Fleet dispatchCoordinate assignments, queues, and production targets
    Control roomMonitor exceptions and enable intervention
    Safety systemApply independent stops, geofences, and fail-safe states

    AI is only one part of the safety architecture. Language models should not control vehicle motion. They can support maintenance knowledge, explain fleet exceptions, summarize shift events, and help controllers retrieve procedures.

    Mixed-fleet support creates bespoke value because each truck model and site has unique dynamics. The enterprise's road network, payload, weather, interaction rules, and incident data shape the models and operating envelope.

    4. How It Reached Production

    Autonomous haulage requires a staged safety case.

    Define the operating domain. Specify roads, speeds, weather, visibility, traffic participants, and failure conditions where autonomous operation is permitted.

    Separate safety and optimization. Production optimization cannot override collision avoidance, emergency stop, or geofence controls.

    Validate mixed-fleet behavior. Test interaction across vehicle sizes, braking profiles, human-driven equipment, light vehicles, and road users.

    Train the control organization. Roles move from driving to supervision, maintenance, remote assistance, and exception response.

    Measure normalized economics. Cost per tonne should include capital, retrofit, network, control room, maintenance, energy, labor, availability, and production.

    Every disengagement, emergency stop, localization failure, and near miss should feed a site-owned evaluation and incident process.

    5. What Mining Leaders Should Take Away

    The Shougang and EACON case demonstrates the scale of value possible when AI is connected to physical operations. It also demands the highest level of evidence discipline because safety and capital intensity are substantial.

    The bespoke layers around mine autonomy and dispatch include fleet optimization, energy models, fuel intelligence, predictive maintenance, field speech, incident RAG, and control-room agents with limited authority. Physical vehicle control should remain inside validated safety systems.

    The core metric is safe total cost per tonne, with energy, availability, production, disengagements, and human exposure reported separately. Large outcome claims are useful signals, but local validation is non-negotiable.

    Sources