Mining has always been a business of moving rock, but a new layer of value lies in the software that decides which rock moves, when it moves, and which machine carries it. Field deployments of autonomous mining systems across more than 30 mines have already handled over 30 million tons of material. That scale marks a significant evolution in operational technology.
Image Credit: Sandwish Studio/Shutterstock.com
The Mining Shift
Early systems described conditions and predicted outcomes for a human planner to act on. Newer systems select actions, negotiate with other machines, and revise plans while a shift is still running, moving the algorithm from advisor to operator.
The word agentic captures this capacity to act. An agentic system pinpoints a goal, senses its environment, chooses among options, and learns from the results of its own choices. In a mining context, these goals could be tons moved per hour, fuel burned per ton, or minutes of shovel idle time.1,2
Click here to download a PDF copy of this page
Dispatching as the First Real Test
Truck dispatching offers the clearest proving ground for agentic AI in the open pit. Conventional mathematical dispatching methods struggle with the autonomy and constant change that modern operations demand, since queues form, roads degrade, and equipment fails on no fixed schedule.2
Reinforcement learning reframes each truck as an agent that learns from repeated interaction with a stochastic mine. The environment covers shovels, destinations, roads, people, and the other trucks competing for the same loading equipment. Assignments then emerge from a learned policy that improves with experience, replacing fixed rule tables written before the pit took its current shape.2
The research record shows how young this field remains. A recent Mining Technology review of reinforcement learning fleet management systems identified 29 desirable dispatching features and found that 17 of them, about 60%, went unaddressed. Every system treated dispatching as a single-objective problem.2
Where Current Algorithms Fall Short
The gaps identified are significant because mines have to balance several conflicting objectives simultaneously. For instance, production targets must be weighed against fuel consumption, tire wear, blending requirements, and road congestion. Currently, the systems in use rely on model-free methods, with the double deep Q network being the most advanced algorithm applied.2
However, multi-objective multi-agent reinforcement learning remains largely unexplored in this setting, even though the dispatching problem naturally involves many learners with competing goals. Recommended paths forward include dueling architectures, actor-critic policy methods, and mixed model-free and model-based designs suited to the structure of haulage decisions.2
State representation design is equally important. Reviewed systems encoded up to 16 attribute types, and one published state vector reached 102 elements. What an agent observes shapes what it can learn, making state representation a design decision with direct consequences for production and fuel spent per ton hauled.2
Digital Twins as the Training Ground
Learning agents need a safe environment to fail, and the mine itself is a costly classroom. A deployed framework pairs each physical operation with a self-evolving digital twin that simulates real scenarios, generates new ones, and supports low-cost training and testing of the complete autonomous system.1
The architecture consists of three main components:
- A model-driven engine encodes the terrain, sensors, and machine dynamics.1
- A data-driven twin replays site telemetry and control commands to refine these models.1
- Parallel computational experiments explore a scenario library to identify better operating policies before they are implemented in the pit.1
Results from cold-weather operations indicate that this approach remains effective even under stress. Autonomous operations continued at temperatures near -42 °C, navigating snow, fog, dust, and low light, while covering over 4.1 million km without a major accident and maintaining round-the-clock production.1
Prediction Moving Toward Autonomous Response
Safety analytics follow a parallel path from measurement toward action.
A recent MDPI Sensors analysis of coal mine disaster prediction traces a progression from traditional machine learning and deep learning to large language models, and then to autonomous agents and multi-agent systems built for hazard warning across gas outbursts, fires, water inflows, roof failures, and dust events.3
China’s record of 168 coal mine accidents in 2022, resulting in 245 fatalities, underscores the urgency of these developments.
More than a third of its mines are classified as highly prone to gas or outbursts. Gas concentration warning systems using recurrent models and stream processing have produced alerts in under eight seconds.3
However, current systems are limited by fragmented data, lack of solid mechanistic grounding, and delayed warnings. The analysis argues for unified intelligent platforms that fuse multiple data types across hazards and scales. This approach aligns well with the coordination capabilities of agent-based architectures.3
The Human Layer Around the Algorithm
Technology adoption in mining is fundamentally linked to the people involved.
A Mining Technology report identified workload, cognitive load, overreliance, trust, and human-centered design as recurring individual concerns. Training and evolving role definitions are also vital factors.4
At the organizational level, technology adoption affects task allocation, safety performance, costs, and company culture. Community impacts, on the other hand, focus on employment opportunities and relationships with Indigenous communities. The report also found that unquantified community impacts weaken a company's ability to meet its corporate social responsibility commitments.4
Situation awareness deserves particular attention as decision authority shifts toward software. Operators who monitor screens instead of driving trucks develop different mental models of the pit, and the quality of those models determines how quickly a person notices an algorithm behaving badly. Effective communication practices and team structures are equally important in this context.4
Plumbing Before Intelligence
The binding constraints sit below the algorithm. Interoperability between vendor systems and inadequate wireless coverage rank as the leading barriers to mining automation, followed by real-time positioning accuracy and cyber safety across connected fleets.4
Each constraint directly affects agentic operation. An agent starved of position data acts on a stale picture of the pit. A fleet split across incompatible control systems cannot negotiate shared haul roads. A breached network turns coordinated autonomy into coordinated risk, since one compromised channel reaches every machine listening on it.4
The realistic near-term picture is narrow agency inside well-instrumented zones, widening as connectivity, data standards, and socio-technical practice mature. Progress will depend on treating mine automation as a combined technical and organizational program, with attention to skills, standards, and community outcomes alongside the algorithm itself.4
References and Further Reading
- Chen, L. et al. (2024). Autonomous mining through cooperative driving and operations enabled by parallel intelligence. Communications Engineering, 3(1), 75. DOI:10.1038/s44172-024-00220-5. [https://www.nature.com/articles/s44172-024-00220-5](https://www.nature.com/articles/s44172-024-00220-5)
- Hazrathosseini, A., & Afrapoli, A. M. (2024). Transition to intelligent fleet management systems in open pit mines: A critical review on application of reinforcement-learning-based systems. Mining Technology. DOI:10.1177/25726668231222998. [https://journals.sagepub.com/doi/10.1177/25726668231222998](https://journals.sagepub.com/doi/10.1177/25726668231222998)
- Lu, P. et al. (2025). Application of Artificial Intelligence in Predicting Coal Mine Disaster Risks: A Review. Sensors, 25(21). DOI:10.3390/s25216586. [https://www.mdpi.com/1424-8220/25/21/6586](https://www.mdpi.com/1424-8220/25/21/6586)
- Codoceo-Contreras, L. et al. (2024). Exploring the impacts of automation in the mining industry: A systematic review using natural language processing. Mining Technology. DOI:10.1177/25726668241270486. [https://journals.sagepub.com/doi/10.1177/25726668241270486](https://journals.sagepub.com/doi/10.1177/25726668241270486)
Disclaimer: The views expressed here are those of the author expressed in their private capacity and do not necessarily represent the views of AZoM.com Limited T/A AZoNetwork the owner and operator of this website. This disclaimer forms part of the Terms and conditions of use of this website.