Source-linked AI summary
Managing Context and Communication in Distributed Agentic UAV Swarms
Andrea Iannoli, Ivan Zyrianoff, Angelo Trotta, Lorenzo Gigli, Marco Di Felice
TL;DR
Distributed UAV agents must manage long-running histories and peer messages without relying on a central coordinator. The paper combines local SLM control with bounded note-based memory and deterministic, interest-aware gossip. In simulated search-and-rescue missions, the policy completed all runs and reduced communication and inference overhead relative to unrestricted flooding.
Problem
Distributed UAV agents operate with partial observations and resource constraints, while long histories and indiscriminate dissemination challenge context and communication management.
Source: §1 · p. 2, §3.1 · p. 4Method
Each UAV runs an independent SLM agent with bounded core, local, and peer-specific note memory and deterministic gossip that uses recipient-specific semantic novelty and recency.
Source: §1 · p. 2Results
The proposed policy completed all runs; Full Flood completed 70–85% and SLM forwarding none, while the proposed policy used roughly half the tokens and sent below 0.5 MB versus 6–7 MB.
Source: §1 · p. 3, §6.2 · pp. 9–10Takeaways & Limitations
In the evaluated search-and-rescue scenario, the proposed policy outperformed the evaluated SLM-based baselines in mission completion, communication overhead, inference-token consumption, and survivor-count accuracy.
Source: §7 · p. 11The evaluation compares information-management strategies under shared lifecycle, prompt, MCP, and safety settings, and the system model assumes local partial views rather than a globally synchronized swarm view.
Source: §3.1 · p. 4, §6.1 · p. 9
Abstract
from arXiv · showhide
Unmanned aerial vehicle (UAV) swarms increasingly rely on language-model agents to provide adaptive mission-level reasoning in uncertain environments. Fully distributed control, in which each UAV hosts an independent Small Language Model (SLM), removes reliance on a centralized coordinator but introduces an information-management problem: long-running interaction histories can degrade the reasoning context, while indiscriminate information dissemination increases communication and inference overhead. We address these challenges with a distributed UAV-agent architecture that enables continuous local SLM control through an event-driven reason-act-observe lifecycle. Runtime knowledge is represented as structured atomic notes and organized into core, local, and peer-specific memory. A deterministic interest-aware gossip engine selectively disseminates these notes according to recipient-specific semantic novelty and recency. We evaluate the architecture using ten UAVs in a simulated search-and-rescue mission. Our approach completes all experimental runs, whereas unrestricted flooding messages completes only 70-85\%, and delegating forwarding decisions to the SLM prevents mission completion in every run. Compared with unrestricted flooding, our approach approximately halves inference-token consumption, reduces transmitted data, and achieves lower survivor-count error.
1 Introduction
The introduction frames distributed SLM control as a way to avoid centralized coordination, but highlights the challenge of managing long-running local knowledge and peer communication. It proposes bounded memory and deterministic, interest-aware dissemination, then evaluates the approach in a search-and-rescue mission.
Introduction
- Centralized language-level control depends on reliable connectivity and can create communication and computation bottlenecks, while decentralized agents face partial views and tight onboard resource constraints.The introduction identifies memory, computation, energy, and latency constraints as reasons to favor smaller SLMs.Source: §1 · p. 2
- Long transcripts can bury mission-critical information, while indiscriminate flooding wastes bandwidth and SLM-based forwarding decisions add reasoning overhead.The paper describes context rot as failure to use relevant information in long inputs, even before the context limit is reached.Source: §1 · p. 2
- Each UAV’s Context Manager stores bounded core, local, and peer-specific memory as atomic notes, supporting continued local reasoning without a central coordinator.The architecture pairs this memory with an independent SLM-based agent and an Interest-Aware Gossip Engine.Source: §1 · pp. 2–3
- The deterministic gossip engine compares notes with each neighbor’s history using semantic similarity and recency, then suppresses, compacts, or forwards them.Forwarding decisions do not require additional SLM invocations; the SLM is used for compaction when selected.Source: §1 · pp. 2–3
- The proposed policy completed all experimental runs and, versus the most successful SLM baseline, used approximately half the inference tokens, transmitted below 0.5 MB rather than 6–7 MB per mission, and achieved lower survivor-count error.The evaluation used ten UAVs in missions involving 5, 10, or 20 moving survivors.Source: §1 · p. 3
2 Related Works
Related work covers language-model reasoning for embodied control, selective communication among cooperating agents, and gossip-based dissemination. The paper presents its contribution as integrating distributed UAV control, semantic memory, and information dissemination within one architecture.
Related work
- Robotic language agents have supported action selection and collaborative planning, while UAV work spans centralized interpretation, edge assistance, and per-UAV language-model decisions.These systems generally use language models for mission-level reasoning, planning, or navigation, with conventional controllers handling lower-level control.Source: §2 · p. 3
- Selective communication methods seek decision-relevant neighbors or compact messages, and semantic communication considers information value for a task, time, or receiver.The reviewed approaches aim to avoid communication that does not aid coordination.Source: §2 · p. 3
- Gossip research addresses freshness, redundancy, topology, and local forwarding; related approaches use semantic routing, interest-based organization, or packet-reception history.The passage describes these approaches as ways to reduce indiscriminate flooding.Source: §2 · p. 3
- The paper distinguishes its work by integrating fully distributed UAV control, semantic memory, and information dissemination in a unified architecture with an independent SLM on each UAV.The cited research lines are described as addressing these building blocks in isolation.Source: §2 · p. 3
3 System Model and Architecture
The system is mission-agnostic: each UAV hosts an independent agent, and mission behavior is specified through its context. The section describes a partial-view swarm model and an onboard reason–act–observe lifecycle that connects tool use, memory, and peer-note exchange.
3.1 System Model
- Each UAV has an independent SLM agent and local state, observes only part of the environment, and exchanges information with nearby peers without assuming a globally synchronized swarm view.Information may propagate over multiple hops, and no specific topology is assumed.Source: §3.1 · p. 4
3.2 Per-UAV Architecture and System Lifecycle
- At each reasoning step, the SLM selects a mission-level action through a structured MCP tool call; execution feedback returns through a callback and triggers the next step.The cycle continues until mission completion, infeasibility, or explicit termination; asynchronously received peer messages are available at the next reasoning step.Source: §3.2 · pp. 4–5
- The WoT layer maps platform-independent capability interfaces to vehicle operations, while the lifecycle creates atomic notes and evaluates each note’s novelty for each current neighbor.The gossip engine suppresses, fully forwards, or compacts notes; only selected compaction invokes the SLM.Source: §3.2 · p. 5
4 Context and Memory Management
The Context Manager limits context growth by separating persistent mission information from locally acquired and peer-associated knowledge. It stores runtime knowledge as structured atomic notes and assembles a bounded context for each SLM invocation.
4.1 Atomic Note Construction
- Atomic notes represent individual pieces of mission-relevant knowledge, with tool responses converted into notes and received notes stored directly.Direct storage preserves received notes' identifiers and semantic contents.Source: §4.1 · p. 5
- Each atomic note carries an identifier, source UAV, generation or observation time, factual content, summary, keywords, semantic tag, and optional application metadata.Tags are assigned based on the type and outcome of the interaction that produced the note.Source: §4.1 · p. 5
4.2 Memory Organization
- The Context Manager combines static Core Memory, a Local Note Queue, and separate Neighbor Note Queues for each peer.Core Memory contains the agent's operating policy, mission specification, UAV metadata, and identifier; it remains unchanged and is included in every SLM invocation.Source: §4.2 · p. 6
- Local and peer-specific queues have independent token budgets, and oldest notes are removed when an insertion exceeds a queue's budget.At each invocation, the context combines Core Memory with the current local and neighbor queues, keeping its runtime portion bounded.Source: §4.2 · p. 6
5 Interest-Aware Gossip Engine
The Interest-Aware Gossip Engine evaluates each candidate note separately for each one-hop neighbor, estimating whether that peer likely already has similar, recent information. It uses the estimate to suppress, compact, or fully forward mission-relevant notes.
5 Interest-Aware Gossip Engine
- The engine compares a candidate note with peer-history notes of the same semantic tag, forwarding only categories designated mission-relevant before execution.Note embeddings are generated locally from concatenated semantic fields; the embeddings are cached and are not transmitted with the note.Source: §5 · pp. 6–7
- Forwarding is decided independently for each one-hop neighbor using that peer's locally maintained history as an estimate of what the neighbor knows.Duplicate notes are identified by note identifier and discarded; forwarding metadata tracks the path and time-to-live separately from note content.Source: §5 · p. 6
- Tag-specific mean embeddings reduce recurrent category-level similarity, while the original embeddings are used when fewer than five matching local notes are available.The method then uses nonnegative cosine similarity to keep scores within [0, 1].Source: §5 · p. 7
- Recent, semantically similar history lowers a candidate's novelty contribution, and the engine uses the minimum contribution as the neighbor-specific score.An empty comparison history receives a novelty score of 1; tag-specific thresholds determine whether to forward, compact, or discard.Source: §5 · p. 7
- A DISCARD suppresses transmission only to the evaluated neighbor, a COMPACT sends an SLM-summarized representation, and a FULL sends the complete note.Discarding does not remove the note from local memory.Source: §5 · p. 7
6 Performance Analysis
The evaluation tests a distributed UAV-agent architecture in a simulated search-and-rescue mission. The proposed policy completed all runs, with lower overhead and survivor-count error than the evaluated SLM-based baselines.
6.1 Experimental Setup
- The case study used ten simulated UAVs to search for 5, 10, or 20 moving survivors, whose movement across strips required the swarm to reconcile repeated detections.Mission completion required all UAVs to finish scanning, recover, land, and disarm.Source: §6 · p. 8, §6.1 · p. 8
6.2 Results
- All runs completed under the proposed policy; Full Flood completed 70–85%, while SLM forwarding completed none, and the proposed policy used approximately half the tokens and below 0.5 MB versus 6–7 MB for Full Flood.The communication comparison is per mission.Source: Table 1 · p. 9, §6.2 · pp. 9–10, Figure 4 · p. 11
- Full Flood had slightly higher survivor-count error than the proposed policy in every scenario, while the Deterministic baseline overcounted as survivors moved across search strips.Proposed-policy errors mainly involved merging visually similar people or counting repeat observations separately; these were less frequent than with Full Flood.Source: §6.2 · pp. 9–10, Figure 2 · p. 10
- Full Flood produced the longest missions; runtime and coordination time included context assembly, forwarding decisions, MCP scheduling, and duplicate, path, and TTL checks.The reported gain was not only from lower network traffic, but also from avoiding the execution-time cost of managing unnecessary information.Source: Figure 3 · p. 10, §6.2 · p. 10
7 Conclusion
The paper presents a fully distributed SLM-based UAV architecture combining bounded structured memory with deterministic, recipient-specific information dissemination. In a simulated search-and-rescue scenario, the proposed policy outperformed the evaluated SLM-based baselines across mission completion, communication overhead, inference-token consumption, and survivor-count accuracy.
7 Conclusion
- The proposed architecture combines a bounded Context Manager with atomic notes and an Interest-Aware Gossip Engine that disseminates notes according to recipient-specific semantic novelty and recency.Future work will evaluate larger and more heterogeneous swarms, other embedded models, and additional mission types.Source: §7 · p. 11