Multi-Agent Reinforcement Learning (MARL) for "Open Autonomy" in Mixed-Fleet Underground Mining

Abstract: Modern underground mining operations are progressively integrating autonomous haulage systems (AHS) and automated Load-Haul-Dump (LHD) vehicles. However, current commercial autonomy architectures are "closed systems" that rely on rigid, centralized dispatch servers and strict zone isolation, requiring human-driven utility vehicles to completely clear an area before automated fleets can operate. This paper reviews the emerging frontier of Multi-Agent Reinforcement Learning (MARL) designed to achieve "Open Autonomy"—a decentralized paradigm where heterogeneous, mixed fleets of autonomous and human-operated vehicles dynamically co-exist and cooperate. We evaluate MARL framework configurations (e.g., Centralized Training with Decentralized Execution [CTDE]), multi- agent communication topologies under intermittent subsurface wireless connectivity, and the integration of game- theoretic collision avoidance algorithms within tight, single-lane subterranean environments.