Distributed Adaptive Multi-Drone Coordination At Scale

Designing and understanding the multi-drone operations is a grand challenge. This is particularly prominent in intelligent transportation systems where swarms of cooperative drones are used for traffic monitoring and last-mile delivery. Although significant technological breakthroughs have been achieved in the control and communication of individual drones, coordinating multiple drones for distributed task allocation remains an open research problem. This involves determining which drones should visit which points of interest on the map, and when, to execute tasks, such as collecting sensing data and inspecting infrastructure, thereby maximizing mission performance (e.g., high completion rate, accurate sensing). Figuring out the best way to allocate these tasks is complex and falls into a class of problems known as NP-hard combinatorial optimization.

Previous work addresses this problem by employing distributed task allocation algorithms, from market-based methods to swarm intelligence. However, such approach comes with three key limitations: (1) Poor scalability in large-scale operations: Existing approaches struggle with long-term, large-scale planning due to high computational and communication costs. (2) Poor adaptability to diverse real-world conditions: Current models often overlook or over-simplify real-world factors, such as drone weight, payload and recharging locations, making it difficult to estimate energy consumption and adapt to environmental dynamics. (3) Costly, risky and oversimplified prototyping: Even scalable and adaptive models need real-world testing with low-cost, safe but not oversimplified prototyping to disentangle the complexity of multi-drone coordination before real-world deployment.

To address these challenges, this thesis contributes to propose a distributed multi-agent coordination model where each agent/drone independently determines navigation, tasking and recharging plans to choose from such that system-wide requirements are met. The model optimizes the discrete plan selection by integrating state-of-the-art reinforcement learning, collective learning and exact algorithms, which enhances both scalability and adaptability in evolving environments under critical hard constraints. As a proof of concept, this work focuses on two scenarios: traffic monitoring in urban sensing and last-mile delivery in logistics. Experimental results demonstrate that the proposed method achieves scalable, energy-efficient and accurate task execution in large-scale spatio-temporal scenarios (e.g., coordinating 1,000 drone dispatches for full-day, city-sized missions), while operating under limited drone and energy resources. These findings offer valuable insights for policymakers, system operators, and technology designers aiming to enhance intelligent transportation systems. Potential benefits include reducing carbon emissions from heavy vehicles, alleviating traffic congestion, improving the quality of urgent medical deliveries (e.g., pandemics), and informing the optimal placement of recharging stations or depots.

Furthermore, the development of an indoor testbed ensures low-cost deployment, operational safety, and the applicability of the proposed system. These advantages are validated through hardware experiments, which confirm accurate energy consumption estimation and minimal collision risk. The proposed testbed bridges the gap between complex algorithmic simulations and practical but oversimplified multi-drone implementations before outdoor testing. It provides a replicable and scalable prototyping, paving the way for broader adoption of advanced multi-drone systems.