Joint Communication Scheduling and Velocity Control for Multi-UAV-Assisted Post-Disaster Monitoring: An Attention-Based In-Context Learning Approach

TL;DR

This work tackles the urgent problem of data collection in multi-UAV post-disaster monitoring by jointly optimizing sensor activation schedules and UAV velocities to minimize packet loss. It proposes AIC-VDS, an attention-based In-Context Learning framework that leverages an edge-hosted LLM to reason from compact, pruned sensory state representations and generate near-optimal transmission schedules and velocity plans without retraining. The approach outperforms a multi-agent DQN baseline and a maximum channel gain strategy, demonstrating faster convergence and substantially reduced packet loss (e.g., up to 91% at moderate network sizes) while accounting for battery, queue, channel, and trajectory constraints. The results suggest that integrating attention-based data compression with LLM-driven decision making can enable real-time, scalable UAV swarm coordination in dynamic disaster environments, with practical implications for improving data reliability and response times in tsunami monitoring.

Abstract

Recently, Unmanned Aerial Vehicles (UAVs) are increasingly being investigated to collect sensory data in post-disaster monitoring scenarios, such as tsunamis, where early actions are critical to limit coastal damage. A major challenge is to design the data collection schedules and flight velocities, as unfavorable schedules and velocities can lead to transmission errors and buffer overflows of the ground sensors, ultimately resulting in significant packet loss. Meanwhile, online Deep Reinforcement Learning (DRL) solutions have a complex training process and a mismatch between simulation and reality that does not meet the urgent requirements of tsunami monitoring. Recent advances in Large Language Models (LLMs) offer a compelling alternative. With their strong reasoning and generalization capabilities, LLMs can adapt to new tasks through In-Context Learning (ICL), which enables task adaptation through natural language prompts and example-based guidance without retraining. However, LLM models have input data limitations and thus require customized approaches. In this paper, a joint optimization of data collection schedules and velocities control for multiple UAVs is proposed to minimize data loss. The battery level of the ground sensors, the length of the queues, and the channel conditions, as well as the trajectories of the UAVs, are taken into account. Attention-Based In-Context Learning for Velocity Control and Data Collection Schedule (AIC-VDS) is proposed as an alternative to DRL in emergencies. The simulation results show that the proposed AIC-VDS outperforms both the Deep-Q-Network (DQN) and maximum channel gain baselines.

Figures (5)

• Figure 1: A multi-UAV-assisted tsunami monitoring. Each UAV follows its trajectory and establishes LoS communication to collect sensory data.
• Figure 2: Overall structure of the proposed AIC-VDS. The MUPDM interacts with an edge-hosted LLM via structured prompts. The LLM receives logged environmental data (e.g., queue length, channel condition, and battery level) and generates optimized data collection schedules and velocities. A feedback loop records system performance and controls continuous adjustment to minimize packet loss.
• Figure 3: Communication protocol for the proposed AIC-VDS. The protocol begins with the UAV querying an edge-based LLM for an optimized data collection schedule and velocity. The LLM selects the most relevant sensor and velocity based on the task description. The UAV positions itself optimally, makes contact via a beacon, receives data and status packets, and confirms receipt. This process is repeated for the next scheduled sensor.
• Figure 4: Flowchart of the AIC-VDS algorithm, illustrating the iterative process for generating optimized data collection schedules and velocities control for a UAV swarm. The process involves data collection, an attention mechanism, an LLM-based decision step, and model optimization.
• Figure 5: Attention mechanism for sensor ranking. Sensor features (queue length, battery level, channel condition) are encoded as vectors. An attention module projects these into query (Q), key (K), and value (V) vectors. Attention scores, computed via dot products and softmax normalization, create a weighted context for each sensor. A final linear layer outputs a scalar importance score for each sensor, which dynamically prioritizes the sensors with high urgency (e.g., high queue, low energy) within the global context. The sensors with the highest score are selected for an action.