Abstract
Unmanned aerial vehicle (UAV) swarms are considered a promising technique for next-generation communication networks due to their flexibility, mobility, low cost, and ability to collaboratively provide services. Distributed learning (DL) enables UAV swarms to intelligently provide communication services, multi-directional remote surveillance, and target tracking. This survey introduces popular DL algorithms such as federated learning (FL), multi-agent reinforcement learning (MARL), distributed inference, and split learning, presenting a comprehensive overview of their applications in trajectory design, power control, and wireless resource allocation. Furthermore, it explores state-of-the-art applications like reconfigurable intelligent surfaces (RIS) and semantic communications, while identifying open problems for future research.
Introduction
While single UAVs are versatile, they are fundamentally limited by finite energy, compute power, and communication range. Transitioning to UAV swarms—teams of multiple drones working together—offers superior coverage and reliability. However, managing a swarm presents a "data deluge" problem: edge devices generate massive amounts of data that are too costly in terms of latency and battery to upload to a central cloud.
Traditional centralized Machine Learning (ML) struggles with the dynamic nature of aerial environments and raises privacy concerns. Distributed Learning (DL) addresses these issues by allowing UAVs to process data locally and share only model updates or limited inference results, effectively turning the swarm into a collective "flying brain."
Core Pillars of Distributed Learning in Swarms
The paper categorizes four primary DL pillars tailored for the unique constraints of aerial swarms:
1. Federated Learning (FL)
In FL, UAVs train models on local data and share only model parameters (weights) with a central aggregator, such as a leading UAV. This preserves privacy and significantly reduces bandwidth usage compared to raw data transmission.
2. Multi-Agent Reinforcement Learning (MARL)
MARL focuses on collaborative decision-making. UAVs act as independent agents that learn optimal policies—such as path planning or resource allocation—by interacting with a shared dynamic environment. Frameworks like Centralized Training with Decentralized Execution (CTDE) are often employed to balance global coordination with local autonomy.
3. Distributed Inference (DI)
This involves splitting the execution of a pre-trained model. It includes Cooperative Inference (sharing layers between drones) and Decision Inference (aggregating local "opinions" to reach a final classification), allowing the swarm to perform complex tasks that exceed the capacity of a single unit.
4. Split Learning (SL)
In SL, a model is physically split. A UAV processes the initial layers (client-side) and sends intermediate "smashed data" to a server to complete the computation. This is ideal for drones with extremely limited onboard compute power.
Key Insights and Applications
- The Semantic Shift: Beyond traditional bit-level transmission, Semantic Communications allow swarms to transmit only the "meaning" or "intent" of data, drastically reducing bandwidth requirements in noisy environments.
- RIS-Enabled Swarms: UAVs can carry Reconfigurable Intelligent Surfaces (RIS)—"smart mirrors"—positioned via DL to reflect signals around obstacles, eliminating dead zones in 6G networks.
- Hybrid Architectures: The survey notes that hybrid approaches, such as Hybrid Split and Federated Learning (HSFL), often outperform pure FL when data is non-identically and independently distributed (non-IID) across the swarm.
Quantitative Impact
The synthesis of current research highlights significant performance gains:
- Coverage: DL-based deployment strategies can improve wireless coverage by up to 58%.
- Latency: Distributed Deep Neural Network (DDNN) frameworks can reduce communication costs by more than 20 times compared to raw data offloading.
- Energy: FL approaches for power control have demonstrated a 35% reduction in the communication rounds required for convergence.
- Reliability: In RIS-UAV scenarios, DL-enabled swarms provide 3.25 times more communication resources than satellite-only configurations.
Conclusion and Future Directions
Despite these advancements, several hurdles remain. System Heterogeneity is a major challenge, as "straggler" drones with low battery can stall the learning process. Additionally, the Communication Overhead of sharing large model updates requires new techniques like over-the-air computation. Future research must move beyond simulations toward Real-world Validation in complex weather conditions and develop Native Protocols that integrate sensing, communication, and learning into a unified stack.
Summary updated