LLM-Centric Agentic AI for UAV Swarms

2 minute read

Published:

UAV swarms promise a lot for search and rescue, environmental monitoring, and precision agriculture, but real deployment is held back by limited situational awareness and intermittent connectivity. Standard LLMs don’t close that gap on their own: they lack temporal memory and native tool interaction, and they can’t generate verifiable, real-time control actions in safety-critical settings. Our paper, LLM-Centric Agentic AI for UAV Swarms, proposes LAUS, a system-level architecture that uses agentic AI to address this.

The central design choice is a strict separation of timescales. The inner loop (milliseconds, onboard, deterministic) handles flight control, reactive collision avoidance, and safety validation, and never invokes the LLM. The mid loop (sub-second to seconds) runs deterministic solvers for trajectory optimization, formation control, and resource allocation, exposed to the agent only as callable tools. The outer loop (seconds to minutes, edge-hosted) is the only tier where the LLM operates, selecting objectives, re-parameterizing solver calls, and scheduling sensing. LLM latency and connectivity outages therefore bound mission-level adaptation, not flight-critical control, which continues under the last validated set-points.

Around this core, LAUS has four main components. Context engineering is the system-level layer that decides what the model sees, remembers, and uses across steps, turning a stateless LLM into an agentic one by managing memory, retrieval, context assembly, and tool use. Tool calling serves every stage of the loop: perception (parsing LiDAR, thermal, and GNSS/IMU data into structured observations), reasoning (retrievers over knowledge bases such as 3GPP standards), and action (resource allocation, formation managers, weather predictors), with outputs gated through independent safety validators before reaching the flight controller.

Agentic memory stores experiences as atomic, interconnected notes with embeddings, rather than rigid predefined schemas. New experiences are encoded as notes, linked to similar past situations, and updated through semantic association, which helps agents keep making decisions during communication blackouts using locally cached, historically grounded context. Adaptation across missions happens without any parametric updates to the LLM: what changes over time is the assembled context, not the model weights.

Finally, a governance layer with human-in-the-loop oversight provides enforceable rules, audit logs for traceability, mandatory human approval for critical decisions, and a fail-safe kill switch, tracked through KPIs such as communication reliability, collision avoidance, energy efficiency, and mission success rate. One note on scope: the paper specifies memory, governance, and oversight as architectural mechanisms; they are not independently instantiated or empirically evaluated.