Safe Reinforcement Learning Methods for Military Applications

Abstract

Reinforcement learning (RL) has shown strong potential for autonomous decision-making in complex and dynamic environments. However, its direct application to military systems raises critical safety, reliability, accountability, and verification concerns. Unlike civilian domains, military applications require strict compliance with operational constraints, human oversight, mission assurance, and rules of engagement. This paper reviews safe reinforcement learning (safe RL) methods for military applications and presents a compact taxonomy of major approaches, including constrained RL, shielding, controltheoretic safety, risk-sensitive RL, offline safe RL, and safe multi-agent RL. The paper further discusses how these methods can be applied to military command and control, autonomous mobility, intelligence, surveillance, and reconnaissance, cyber defense, and human-machine teaming. Finally, we argue that safe RL for military systems should be treated not only as an algorithmic optimization problem, but also as an assuranceoriented framework integrating constraints, verification, runtime monitoring, and human control.