Existing Gaps In Reinforcement Learning For Drone Warfare

Pascal Leroy, Arthur Louette, Damien Ernst

Abstract: Drones have changed warfare and are deployed daily on the battlefield for surveillance or as offensive and defensive weapons. While humans continue to control drones and weapon systems, the transition to autonomous control, which removes the human decision, is imminent. Indeed, advances in artificial intelligence (AI) are extremely rapid and AI-driven drones seem to represent the future of warfare. This motivates the need to improve systems to face autonomous drones and build better ones. Reinforcement learning (RL) is a paradigm of AI focusing on the resolution of sequential decision-making problems. Its deployment in robotics shows its potential to address complex real-world challenges. After presenting RL foundations with a practical battlefield example, we propose a framework to deploy RL in robotics. We identify five axes of complexity to deploy RL on robots for any real-world problem. These axes allow us to analyze the state-of-the-art and identify gaps required by the future of drone warfare. We conclude the paper with a roadmap to bridge these gaps and ethical considerations.