Abstract
As Unmanned Aerial Vehicle (UAV) swarms advance toward higher levels of autonomy, the communication overhead of high-level semantic instructions becomes a critical bottleneck. This paper proposes a novel pipeline using Large Language Models (LLMs) as "semantic compression engines." By summarizing complex mission instructions into compact, intent-rich formats, the system significantly reduces bandwidth requirements while maintaining high task success rates in constrained, multi-hop environments.
Introduction
Modern UAV swarms increasingly leverage Foundation Models (FMs) for complex reasoning and spatial understanding. However, communicating high-level semantic context (e.g., "Search the north quadrant for survivors and regroup at point B") over low-bandwidth links leads to "instruction bloat" and latency. Traditional protocols like MAVLink are insufficient for these high-level semantic needs. This research investigates whether an LLM can autonomously compress its own instructions to save bandwidth without losing the mission's core meaning.
Methodology: The Semantic Compression Pipeline
The authors introduce a communication-execution pipeline where an LLM serves as the central intelligence for both task planning and message compression. The architecture follows a hierarchical structure:
- Command Node (UAV-1): Receives detailed mission prompts and uses an LLM to generate a "zipped" semantic format ($M_{zip}$) that retains critical intent while stripping redundancy.
- Relay Nodes (UAV-2/3): Efficiently forward these compressed instructions across the swarm.
- Executor Nodes: Receive the compressed $M_{zip}$ and use local LLM context to parse and execute the task.
Experimental Setup
The framework was evaluated across four 2D simulation scenarios—Simple, Standard, Complex, and Extreme—with swarms of up to 8 UAVs and communication constraints (7.0 grid unit radius). Performance was measured using four key metrics:
- Compression Ratio (CR): The ratio of compressed to original text length (lower is better).
- Semantic Preservation (SP): Meaning retention measured via BERTScore and DeBERTa-v3.
- Bandwidth Utilization (BU): Real-world efficiency under a simulated 1 Mbps link.
- Success Rate (SR): The percentage of UAVs successfully reaching the target area.
Key Results and Insights
The study evaluated nine mainstream LLMs, revealing a critical trade-off between compression efficiency and mission success.
The Semantic Fidelity Threshold
A strong positive correlation exists between Semantic Preservation (SP) and Success Rate (SR). Models that compress too aggressively, such as LLaMA-4-Maverick-17B (CR: 0.241), often sacrifice critical mission details, leading to the lowest success rate (0.641). Conversely, models like GPT-4o maintain a stable balance, ensuring high success rates through moderate compression.
Reasoning vs. Linguistic Capability
Interestingly, reasoning-specialized models like DeepSeek-R1 and GPT-o3 did not significantly outperform standard models in this specific task. This suggests that general linguistic summarization capability is currently the primary driver for effective semantic compression in swarm coordination, rather than deep chain-of-thought reasoning.
Impact of Complexity
As environmental complexity and swarm size increase, maintaining instruction consistency becomes more challenging. In "Extreme" scenarios, the cumulative effect of multi-hop relays and interaction density leads to a noticeable decline in both semantic preservation and physical success rates.
Conclusion and Future Work
This research demonstrates that LLMs can effectively serve as semantic compression engines, reducing instruction sizes by 30% to 75%. While current results are promising in 2D simulations, future work must address 3D spatial dynamics, physical layer noise (e.g., packet loss), and the integration of multimodal data (visual compression) to further enhance autonomous swarm intelligence.
Summary updated