Multi-Agent Machine Learning for Coordinated Drone Navigation

Abstract

Coordinated navigation of many drones as a large team is a requirement for more and more real-world missions, such as disaster recovery, search-and-rescue, surveillance, and delivery. Coordinating the collective behavior in a team of drones is not trivial. Drones have to constantly respond to dynamic environments, avoid collisions, share constrained airspace, while balancing autonomy and cooperation. In this paper, we present a multi-agent machine learning approach to coordinated navigation of teams of drones that is an alternative to current centralized control approaches. Each drone models, as an intelligent agent, learns from interactions with its environment and other drones. Using multi-agent reinforcement learning (MARL), we include real-time drone control to successfully find collision-free path planning, dynamically assign roles to execute missions, and make energy-aware decisions so that robots can work together in real-time. Unlike current rule-based systems, our learning-based framework can learn in new scenarios, tolerate communication delays and scale without training for size of the swarm. Through simulations in both urban and unstructured terrains, our research shows that we can enable emergent cooperative behaviors in teams of drones to learn to not only navigate, but to negotiate airspace and fit together in ways that better accomplish a mission combined than would be produced per drone.