FROM RPA TO AGENTIC AI: A NEURO-SYMBOLIC ARCHITECTURE FOR NATURAL LANGUAGE CONTROL OF LEGACY MINING SOFTWARE Maycol Jhonatan Benavides Sánchez Department of Mining Engineering, GEA AI Research Lab, Peru (benavidesmaycol81@gmail.com) ABSTRACT The mining industry faces a persistent interoperability challenge: critical workflows are locked within "legacy" proprietary software suites (e.g., for geological modeling, mine planning, and scheduling) that lack modern Application Programming Interfaces (APIs). Consequently, highly skilled engineers spend a significant portion of their time acting as "human middleware," manually transferring data and executing repetitive click-sequences between disparate systems. While Robotic Process Automation (RPA) has offered a partial solution, traditional macro-based scripts are brittle and incapable of handling unstructured requests. This paper introduces Mining Software Agents (MSA), a next-generation Neuro-Symbolic Architecture that combines the semantic reasoning of Large Language Models (LLMs) with the deterministic safety of Symbolic Planning (STRIPS) to control mining software via Natural Language. Unlike standard RPA, MSA employs a dual-layer cognitive engine. The Semantic Layer translates human intent (e.g., "Update the topography and calculate cut-and-fill") into a structured goal state. The Planning Layer then dynamically generates a valid sequence of actions using a backwardchaining algorithm, ensuring operational safety by verifying preconditions before execution. On the client side, a Computer Vision-driven Agent interacts with the software's Graphical User Interface (GUI), mimicking human perception and haptic input. We present a validation case study demonstrating the system's capability to autonomously navigate logic-gated workflows, proving that Agentic AI can bridge the gap between modern generative intelligence and legacy mining tools without invasive integration. KEYWORDS Agentic AI, Large Language Models (LLM), Neuro-Symbolic AI, Interoperability, Mining 4.0, Computer Vision, STRIPS Planner. 1. INTRODUCTION In the era of Mining 4.0, the value chain is increasingly data-driven, yet the "last mile" of execution remains manual. Mine planning departments rely on specialized software suites (e.g., Vulcan, Datamine, Whittle, Leapfrog) that operate as isolated data silos. These systems are often "closed," meaning they do not expose RESTful APIs or SDKs that allow for external orchestration. This forces engineers to perform manual routines—such as updating topographies, generating haulage routes, or calculating volumes—that are algorithmically deterministic but operationally manual. Traditional integration approaches, such as database coupling or reverse engineering, are often cost-prohibitive or violate software license agreements. Robotic Process Automation (RPA) offers a non-intrusive alternative, but standard RPA tools rely on rigid coordinate-based scripts that break
RkJQdWJsaXNoZXIy MTM0Mzk2