SwarmChain: Collaborative LLM Inference for UAV Swarm Control

Unmanned aerial vehicle (UAV) swarms have the potential to perform sophisticated collaborative tasks. However, efficiently managing UAV swarms in harsh environments without ground command centers remains a significant challenge. The widespread application of large language models (LLMs) offers a promising solution to this challenge. By interacting with LLMs, UAVs can generate fluent and coherent responses in real-world tasks. Nevertheless, deploying LLMs on small UAVs is challenging due to their limited computational power and resources. Additionally, current LLMs may suffer from issues such as content generation uncertainty and hallucinations. To address these issues, we develop SwarmChain, a collaborative LLM inference system for UAV swarm control. We first design a tensor parallelism based collaborative reasoning framework (CoLLM) to enable multi-device collaboration for distributing computational loads. Then we propose a resource-aware adaptive load scheduling algorithm (Also) to achieve load balancing for LLM inference tasks. Finally, SwarmChain leverages LLMs to parse instructions, enabling direct interaction with UAV APIs to ensure that UAVs take actions according to task requirements. Experiments are conducted using StableLM and Llama2 models on GPU(Graphics Processing Unit)-less devices (e.g., Raspberry Pis), and the flight states of UAV swarms are simulated using Airsim. SwarmChain achieves a performance improvement of 1.9– 2.3× compared to Llama.cpp MPI, reduces inference latency by 33.33%–52.38%, and achieves an instruction parsing accuracy of 88.9%. The experimental results demonstrate that SwarmChain significantly enhances the task execution efficiency and stability of UAV swarms in complex environments.