Mixed Fleet Operations in Open-Pit Mines: Full-Scale Deployment of 1,900+ Autonomous Mining Trucks Alongside Manned Vehicles
By: CiDi Inc.
In open-pit mining, mixed-fleet operations reflect the high level of autonomy solution providers can now deliver. A common misconception about Level 4 autonomy in mining — autonomous haulage capable of executing haulage tasks without an operator onboard — is that it is mature because it is conceptually feasible in closed-loop or restricted environments. Most mine sites, however, do not conform to such a model. Autonomous trucks, much like autonomous road vehicles, must often operate simultaneously with manned vehicles on the same site. That capability enables flexible haulage for a diverse set of mining scenarios, from coal to aggregates to metals. Hence substantial engineering and operational complexity separates a conceptual closed-loop application from a working mine site, requiring a new system paradigm altogether.[VP1]
Mixed operations arise from one or more realities:
Mine owners phase in site-wide autonomy gradually, often aligning deployment with new contract sections.
Operators retain existing manned fleets and drivers.
As of 2026, autonomous haulage adoption among mining contractors remains selective.
Specialized equipment — excavators, drills, or loaders — may remain manually operated, requiring seamless interaction with autonomous haulage trucks.
In these cases, operators adopting open-pit autonomy quickly learn that the challenge goes beyond simply moving an autonomous truck from point A to B. Comprehensive capabilities require traffic management and right-of-way arbitration, coordinated across a shared production network. Trucks must share haul roads, intersections, loading areas, and dump points with manned vehicles and other equipment without reducing production throughput. In large coal and aggregates operations, mixed fleets can number in the hundreds. Without this capability, autonomous trucks remain confined to highly geofenced or restricted areas, limiting deployment.
Across multiple open-pit sites, this account draws on several years of fleet and technology deployments that enable continuous, production-scale autonomous operations in mixed environments. More than 1,900 autonomous trucks operate around the clock alongside manned trucks, excavators, and support equipment in 40 mining sites worldwide. These include regular production and haulage for some of the largest coal operations in northwestern China, as well as at a diverse set of aggregates and metals operations.
To achieve efficient mixed operations, we find that required capabilities go beyond quantifying sensor suites, compute benchmarks, and discrete autonomy features. Operating principles and the scenario ontology that matters in mining come into focus when a system has to meet production targets shift after shift.
The First Rule of Mixed-Fleet Deployment: Matching or Exceeding Human-Driven Fleet Efficiency
Autonomy programs understandably treat mixed-fleet operations as phased technology migration. An open-pit mining site or quarry first establishes a segregated autonomous zone and deploys trucks within it. After validating stable operation then comes incremental expansion.
In practice, most mines cannot suspend production for technical validation. They require production to continue throughout deployment. Autonomous systems must therefore integrate into existing production systems rather than the other way around. The central design requirement is dynamic priority that incorporates environmental intelligence and mission-based centralized dispatching, rather than operation in closed-loop isolation.
Another trap is that ad-hoc traffic rules that force manned vehicles to accommodate autonomous ones can also produce unintended consequences. At one coal mine during deployment, we tested autonomous-only operating zones and required manned vehicles to slow and yield before entry. Segregated zones and mandatory yielding created moving bottlenecks and reduced fleet efficiency by 5-10%. Although conservative from a traffic-management perspective, these arrangements often degrade operational results.
The autonomous system must instead infer the operating intent and production cadence of surrounding equipment through edge-based intelligence and dynamic priority. Rather than enforcing a fixed right-of-way hierarchy, dynamic priority continuously arbitrates traffic using real-time production context—load status, route, cycle time, queue position, and machine role. Operators that begin site-wide autonomy with this principle achieve more reliable mixed-fleet performance.
The Environmental Intelligence Stack for Autonomous Fleets
To implement dynamic priority, the single vehicle intelligence stack must support flexible decision making and replanning at the edge. This rests on three layers: perception, prediction, and explainable planning. The differentiator is that each layer needs to support flexible and continuous production, not merely collision avoidance.
Perception Stack. In our deployments, a reactive Autonomous Haulage System (AHS) perception system has required roughly 120 m of reliable detection on truck-sized objects and 80 m on personnel or light vehicles — after safety margins — just to execute emergency braking. Dynamic priority is sensitive to unnecessary braking, so the effective look-ahead requirement is even greater.
Despite 30-40km/h operating speeds, compute needs remain high at the edge, considering haul trucks have large geometric blind spots and spend significant time queuing or reversing. Avoidable brake events can disrupt traffic flow, leading to downstream equipment idling and reduced production.
Besides just certifying a collision-free envelope, perception must detect obstacles with enough immediacy to support smooth, flexible re-planning maneuvers on the fly, such as overtaking that manifests as subtle lateral vehicle steering movements. This requirement favors environmentally robust multimodal perception: overlapping LiDAR, radar, and camera coverage fused into a 360-degree representation.
Prediction. Autonomous trucks need to predict human driver behavior when sharing the road. A unified dispatch system lets autonomous fleets train on both sensor and dispatch data, each of which carries information about human operating behavior. At CiDi, our team developed inverse reinforcement learning methods to iteratively learn robust reward models from a limited number of mixed-scenario demonstrations. The prediction module must answer: where does the manned vehicle I see intend to go? Is it moving to queue? Is it loaded and traveling uphill? Correct answers directly affect safety and efficiency.
Explainable planning. Although many dynamic-priority functions can be trained end to end, mining operators also require explainability from autonomous fleet behavior — especially explainability that matches their physical understanding of the site. Planned speed is one example, especially in relation to mining road design. Unlike public roads, mining roads are unstructured. Grade, width, and purpose vary, and roads are designed and even evolve according to mining requirements. To navigate road features—such as intersections, roundabouts, merge points, and one-way passages — autonomous trucks need to build representations of navigable spaces without human intervention. Especially where visibility is limited, this helps the vehicle determine efficient trajectories with sufficient safety margins.
Overall, we believed early on that vehicles need to be free to independently execute haulage tasks as well as mission-critical safety maneuvers, in ways that take human-driven vehicles into account.
Dynamic Priority via Centralized Dispatch
Built on top of the vehicle intelligence foundation, we then developed centralized dispatch that projects the mixed fleet into a coordinated scheduling domain, one in which right-of-way emerges from a rolling optimization, rather than a set of fixed hierarchies.
While manned vehicles cannot always be centrally dispatched, sufficiently coordinated driverless fleets eventually determine the efficient flow of traffic. This gives human dispatchers more fine-tuned control over mining traffic.
Hence, one can treat the automated dispatch system as the main mechanism servicing the mining operation’s own haulage needs. The main system itself takes the universe of high-frequency inputs, including fused perception, vehicle state, route topology, destinations, queue occupancy, loading and dumping status, and predicted trajectories. It then repeatedly solves a constrained multi-agent problem in which safety and geofence limits are hard constraints, while delay, queue growth, energy use, and throughput form the multi-criteria objective.
Because mining roads are unstructured, the system recomputes optimal passing order as traffic conditions and mining requirements change; an intersection that favors a loaded truck on one cycle may hold that same truck the next time around, if its downstream path is blocked. This is similar in concept to the “replanning” problem familiar to individual autonomous vehicle systems, where at any given moment, vehicles constantly reconsider the reasonable path in time-space coordinates. Here, this problem expands into the multi-vehicle domain.
At intersections, the automated dispatcher evaluates each traffic encounter against the objectives. It weighs fleet motion, load state, and queue lengths, issuing a context-specific, coordinated passing order that avoids deadlock, blocking, and moving bottlenecks. The system may hold an empty truck to protect a loaded truck’s exit path, while giving the loaded truck route priority even if it arrives later. Free space is maximized overall for flexible entry and exit. This optimization runs continuously, adjusting as vehicles enter and leave intersections.
Dump areas with unloading coordination highlight how this pipeline depends heavily on cloud-edge fusion. Each autonomous truck detects nearby manned trucks and uploads their position and predicted trajectory. Because several trucks may observe the same manned vehicle with slight differences in contour and position, the cloud performs track association and deduplication before classifying the manned truck’s state. Classifications — such as loaded, queuing, approaching a queue, or preparing a dump — determine right-of-way for manned and autonomous vehicles, and they allow the dispatcher to sequence entries into the dump without forcing manned operators to yield according to an artificial rule set.
Software Architectural Decisions
While the above requirements for fleet and cloud would suffice for a basic AHS system, there are performance gains from even deeper architectural decisions.
We noticed early that mining autonomy is often framed as a choice between cloud and edge. Since mixed operations need both real-time reactivity and organized cooperativity, in an ideal architecture both should sit in perfect harmony.
Here, the system paradigm required is different than the one posed by autonomy in other verticals such as highway trucking; mining demands a multi-agent framework where the decision and planning software layer sits across the entire vehicle fleet, able to orient itself towards collective dispatch goals and dynamic priority, with the mixed fleet depending on the two control layers — vehicle and cloud — working in concert with clearly defined functional boundaries.
Manned-Vehicle V2X as an Added Layer of Redundancy
An added layer of redundancy comes from extending Vehicle-to-Everything (V2X), especially Vehicle-to-Vehicle (V2V) to manned vehicles via retrofit onboard units (OBU). Among autonomous trucks, V2V forms a mesh for point-to-point negotiation and even stronger information and intention sharing. Each truck broadcasts its state, intent, and trajectory. Fitting manned vehicles with a retrofit V2V onboard unit adds a separate tier that provides added signal and information.
The OBU sees beneficial use on the driverless fleet, but with the addition of the V2V on manned vehicles, the system broadcasts vehicle positions, speeds — and where available, operator intents. Autonomous trucks receive these broadcasts directly, independent of cloud connectivity, and fold the received information into prediction and right-of-way models. Manned vehicles then function as active participants in the V2V network rather than passive obstacles detected only by onboard sensors. Since V2V allows for direct wireless communication, this tier allows for added safety and efficiency unencumbered by environmental disturbances such as dust, sensor failure, and even obstruction.
This tier supplements perception rather than replacing it, and the design has to assume it will be incomplete. Light vehicles, contractor traffic, and visiting vehicles are frequently unequipped, and an OBU can fail or be switched off, so the autonomous fleet must continue to handle every manned vehicle as though it carried no unit at all.
Not every mine permits this added tier, and the retrofit carries its own fitment and maintenance burden. For example, autonomous trucks can negotiate with human drivers through V2V, with the manned operator receiving alerts or yield requests through the OBU interface.
Operator Workflows, Field Feedback, and Training
In our experience, mining staff meet autonomy with a mixture of curiosity and initial resistance. Initially, we observed an increased task load for human operators and dispatchers. Early-stage deployments require excavator operators, for example, to enter material types and via an onboard terminal, as well as to direct autonomous trucks to stop at preferred loading positions. These additional steps disrupted familiar routines. Human-driven truck operators also faced a learning curve, particularly in interpreting autonomous vehicle behavior and yielding decisions. V2V alleviates part of this: when an autonomous truck detects a disabled vehicle beyond a blind bend, a timely warning gives the approaching driver information before the hazard comes into view.
Gains come when field-based feedback is able to shape the dispatch system’s response. Truck allocations now adjust to each excavator’s observed loading rate, reducing machinery idling. The system also uses historical haulage records, including trips made by human-driven vehicles, to build a mine-wide world model of haulage traffic, allowing the system to forecast congestion and adjust truck movements. More consistent truck arrivals can increase completed loading cycles and provide better data to transparently evaluate excavator operators. Hence, communicating system productivity improvements over time is an important part of promoting staff acceptance.
Adoption also requires systematic training. For any given deployment, project managers or field engineers typically deliver around seven sessions on average over approximately one week, followed by assessments for excavator operators and human-driven truck drivers. Training covers terminal inputs, autonomous vehicle perception and priority rules, V2X warnings, and emergency response. Site managers also spend time reinforcing the most important procedures in daily briefings.
Smooth Transition to Large-Scale Mixed Operations
When human-operated vehicles enter the unified dispatch framework at the outset, a mine can move from manned operations to mixed-fleet operations without an extended phase of segregated haulage. Deployment then proceeds through two stages:
The smoothest deployments often gave frontline dispatchers meaningful override authority. Temporary override preserves the final layer of operational control that a complex production system demands.
Conclusion
Mixed-fleet deployment in open-pit mining has moved from feasibility to production performance, and the focus is shifting from single-vehicle autonomy to system-level operational capability at scale.
The innovations are often less about algorithms and more about systems engineering — the software architecture decisions that need to be made around designing a large-scale robotic fleet.
A production-ready system coordinates autonomous and human-operated equipment under a common dispatch framework. The traffic management model and production targets form the basis of a system that oversees autonomous haulage fleets in execution — in a way that adjusts for human drivers in the mix. Such a system supports predictable vehicle behavior and distributed decision-making under appropriate human oversight. The practical benchmark is straightforward: autonomous trucks and human-operated equipment operate alongside one another while maintaining safety and matching or exceeding existing production efficiency at scale.
