AeroLat: Channel-Aware Latent Space Semantic Communication for Decentralized UAV Swarms

Rajdeep Ghosh; Goparaju Venkata Seshachala Sree Vatsava; Sudip Misra

cs.NISep 15, 2026

AeroLat: Channel-Aware Latent Space Semantic Communication for Decentralized UAV Swarms

Authors: Rajdeep Ghosh, Goparaju Venkata Seshachala Sree Vatsava, Sudip Misra

Organizations: Department of Computer Science and Engineering, Indian Institute of Technology Kharagpur, Kharagpur 721302, India

Abstract

Communication in latent space offers an intriguing alternative to symbolic messages for decentralized autonomous Unmanned Aerial Vehicle (UAV) swarms operating over bandwidth-constrained, time-varying wireless links. However, when homogeneous frozen models are prompted with discretized perceptual inputs, their broadcast states collapse toward the shared prompt template. In view of this, we propose AeroLat, a channel-aware latent semantic communication framework that uses evidence injection. The resulting latent states are then passed through an explicit communication model that encompasses bandwidth-limited serialization, additive noise and information staleness, which facilitates a joint assessment of communication fidelity and swarm-level coordination. Across multi-seed simulations, AeroLat provably remains resilient to codec choice, faults and increasing swarm size. It consistently reproduces the latent-swarm anomaly, while no-whitening controls recover the collapse. In particular, AeroLat is capable of reducing false similarity by 97.5%.

Figures & tables

		latent-space semantics			communication & platform realism
Work	Domain	LLM latent	Anti-	Measured	Channel	Mobility	Physical	Seeded
		payload	collapse	info. bounds	model	/ node speed	platform	stats
Interlat [ 1 ]	latent MAS	✓	✗	✗	✗	✗	✗	p
LatentMAS [ 2 ]	latent MAS	✓	✗	p	✗	✗	✗	✓
CIPHER [ 3 ]	latent MAS	✓	✗	✗	✗	✗	✗	p
ThoughtComm [ 4 ]	latent MAS	✓	p	p	✗	✗	✗	✓

TABLE I: Related works across the two capability clusters this problem requires. ✓ = addressed; ✗ = not addressed; p = partially addressed. The upper block represents latent-space multi-agent communication; the middle block represents semantic communication; the lower block represents UAV networking and mobile computing.

Symbol	Description
$\mathcal{D}_{i}$	agent (drone) $i$
$T_{\mathrm{inf}}$	perception/inference interval ( $3$ s)
$I_{i}(t)$ , $v_{i}(t)$	frame and CLIP image embedding of $\mathcal{D}_{i}$
$p_{i}(t)$ , $H_{i}(t)$	CLIP posterior over $C{=}13$ classes; its entropy
$x_{i}(t)$	evidence-verbalized prompt
$h_{i}(t)$	raw latent from last $k{=}8$ hidden states in $\mathbb{R}^{7168}$

TABLE II: LIST OF SYMBOLS

Fig. 1: AeroLat system architecture for one agent $\mathcal{D}_{i}$ . (A) zero-shot CLIP perception emits the full posterior, image embedding, visual entropy, and pose rather than an $\arg\max$ label; (B) that evidence is verbalized into the prompt of the frozen LLM; (C) template whitening removes the shared common-mode subspace. The centre part depicts the asynchronous TCP latent mesh, one connection per packet with sequence numbers and transmit timestamps so receivers can measure message age. (D) staleness-aware attention fusion; (E) the symbolic flight stack.

Fig. 2: Semantic rate–distortion over the real latent corpus (ten seeds). Dense quantization from fp32 ( $229.5$ kb/state) to int8 ( $57.5$ kb) is semantically lossless ( $D_{\mathrm{sem}}\leq 2\times 10^{-4}$ ); int4 ( $28.8$ kb) costs $D_{\mathrm{sem}}=0.067$ in geometry while receiver probe accuracy remains $0.905$ : within this codec family, rate can be cut $8\times$ before task distortion begins.

Fig. 3: AWGN robustness. From 30 dB down to 0 dB SNR, the geometric distortion grows ( $D_{\mathrm{sem}}\,0.000\!\to\!0.293$ ) while receiver probe accuracy is flat ( $0.892\!\to\!0.889$ ): the whitened latent code degrades gracefully, with no digital cliff.

Operating point	bits/state	$\hat{I}_{\mathrm{NCE}}$	Fano	$\rho$ (b/kb)
fp32, noiseless	229,504	$\geq 2.52$	1.91	0.011
fp16	114,816	$\geq 2.52$	1.91	0.022
int8	57,472	$\geq 2.52$	1.90	0.044
int4	28,800	$\geq 2.52$	1.88	0.088
fp32 @ 10 dB	229,504	$\geq 2.52$	1.83	0.011
int8 @ 10 dB	57,472	$\geq 2.52$	1.83	0.044

TABLE III: Measured information per state and information density

\rho

(neural InfoNCE and Fano bounds; task ceiling

H(Y)=2.52

bits; the marginal overshoot at 0 dB reflects finite-batch estimator variance).

Fig. 4: Task-relevant information (InfoNCE and Fano lower bounds, bars) and density $\rho$ (line) per channel operating point. The state carries the entire task entropy essentially and retains it under int4 quantization and 0 dB AWGN simultaneously.

Run	$\mathcal{L}_{\mathrm{task}}$	$\mathcal{L}_{\mathrm{sep}}$	$\mathcal{L}_{\mathrm{pref}}$	$\mathcal{L}_{\mathrm{geom}}$
full	$2.37\!\to\!0.39$	$-.01\!\to\!-.11$	$3.36\!\to\!1.52$	$.38\!\to\!.07$
no-sep	$\to 0.39$	(off)	$\to 1.55$	$\to.08$
no-pref	$\to\mathbf{0.008}$	$\to-.13$	(off)	$\to.13$
no-geom	$\to 0.41$	$\to-.10$	$\to\mathbf{1.87}$	(off)

TABLE IV: Compressor training and leave-one-loss-out ablations: initial

\to

final loss values (full run: 30 epochs; ablations: 15; for parity the full run reads

\mathcal{L}_{\mathrm{task}}{=}0.405

at epoch 14).

Metric	Argmax prompt	Evidence prompt
	(prior system)	(AeroLat)
Probe accuracy, raw $h$	$0.593\pm 0.071$	$\mathbf{0.923\pm 0.020}$
Probe accuracy, whitened $\psi$	$0.500\pm 0.065$	$\mathbf{0.878\pm 0.037}$
Fano MI (raw), bits	$0.61$	$\mathbf{1.96}$
Off-diag. cosine, raw	$\mathbf{0.9978}$	$0.9989$
Off-diag. cosine, whitened	$0.184$	$\mathbf{0.025}$

TABLE V: The collapse, reproduced and repaired (

n{=}400

per condition; chance

=0.167

). Row 4 is the earlier draft’s flagship “

1.00

similarity,” explained.

Fig. 5: Baseline comparison across twenty seeds. Left to right: relay success rate, coordination latency $\tau$ , information density $\rho$ , and hypothesis breadth $P_{50}$ .

Fig. 6: Consensus under packet loss (i.i.d. and Gilbert–Elliott burst; $N{=}5$ fusion mesh, ten seeds). Final same-scene consensus is $1.00$ at every loss rate up to $50\%$ ; loss costs convergence time only ( $1.9\to 2.3$ rounds).

Fig. 7: Age of information under bandwidth caps ( $50$ ms link latency): p95 message age versus link rate per codec. At $500$ kb/s, quantization moves the latent message from $552$ ms (fp32) to $208$ ms (int8) and $151$ ms (int4), against JSON’s $100$ ms floor; at ${\geq}5$ Mb/s all codecs converge.

Fig. 8: Mesh scaling (ten seeds): rounds to $C\geq 0.9$ grow sub-linearly ( $1.4\to 2.3$ for $N{=}3\to 20$ ), and under a shared $20$ Mb/s medium with int8 payloads and $10\%$ loss the mesh holds $0.90$ delivery with unchanged convergence at every $N$ .

Configuration	Missions	SR	Encode ms (med/p95)	Whitened $C(t)$	Delivery	Mesh delay
Symbolic CoT	2	$0.00$	$3{,}137\,/\,3{,}464$	$0.638\pm 0.083$	$1.00$	$\mathbf{22.6}$ s
JSON dispatch	3	$0.33\pm 0.47$	$71\,/\,79$	$0.886\pm 0.041$	$1.00$	$39$ ms
Latent, dead drone $\{1,2,3\}$	3	$0.67$ ( $2/3$ )	$114$ – $120$	$0.83\pm 0.06$	$1.00$	$13$ – $60$ ms
Latent, $20\%$ loss + int8 + $50/25$ ms jitter	2	$0.50$	$117\,/\,128$	$0.656\pm 0.100$	$0.80$	$136$ ms
Latent, no whitening (control)	2	$0.00$	$146\,/\,159$	$\mathbf{0.999\pm 0.000}$	$1.00$	$134$ ms
Latent	5	$\mathbf{1.00}$	$117\,/\,127$	$0.678\pm 0.094$	$1.00$	$51$ ms

TABLE VI: Eighteen-mission SITL campaign. One mission hang was reclaimed by the per-mission watchdog;

17/18

completed nominally. Bold marks the values discussed in the text: the latent codec’s success rate and the two failure signatures (CoT’s mesh flooding, the no-whitening control’s collapsed consensus).

Fig. 9: The frame of what the drones actually see. The heterogeneity is the point: at any instant the swarm holds genuinely different, individually ambiguous views— $\mathcal{D}_{1}$ reads the spawned cluster as debris at only $p{=}0.25$ with $H{=}2.28$ nats—and it is exactly this residual uncertainty that the whitened latent channel transports and the argmax-based JSON codec destroys .

$N$	Missions	SR	Enc. ms	Whitened $C(t)$	$\operatorname{erank}$	Msg age
3	5	$1.00$	$117$	$0.678\pm 0.094$	$1.9$	$51$ ms
5	3	$1.00$	$119$	$0.649\pm 0.070$	$3.8$	$78$ ms
10	2	$1.00$	$148$	$0.632\pm 0.068$	$7.6$	$2.9$ s
15	$0/2$	watchdog $900$ s / boot failure; $11/15$ airborne, $132$ frames

TABLE VII: Embodied scaling campaign. The campaign is incremental by design: each size must succeed before the next is attempted, so the failure boundary is itself a measurement.

Explore similar work

Managing Context and Communication in Distributed Agentic UAV Swarms
Oct 1, 2026Andrea Iannoli, Ivan Zyrianoff, Angelo Trotta +2Multi-Agent LLM Systems Swarm Robotics
Department of Computer Science and Engineering University of Bologna Bologna, Italy · Technology Innovation Institute (TII) United Arab Emirates
LAEI: Layered Autonomous Edge Intelligence Framework for Robust UAV Swarm Operations
Jun 8, 2026Changmin Park, Wooyong Jung, Hwangnam KimAerial Robotics Swarm Robotics
School of Electrical Engineering Korea University Seoul 02841, Korea
AeroWeaver: An Embodied-Agent Harness for Weaving Aerial Skills into Distributed, Adaptive Swarm Execution
Sep 16, 2026Jiabin Lou, Yirong Yang, Haopeng Wang +6Aerial Robotics Swarm Robotics
Beihang University, 37 Xueyuan Road, Haidian District, Beijing 100191, China

Oct 1, 2026J/K move · Enter open · S save

Managing Context and Communication in Distributed Agentic UAV Swarms

Andrea Iannoli, Ivan Zyrianoff, Angelo Trotta, Lorenzo Gigli+1

Department of Computer Science and Engineering University of Bologna Bologna, Italy · Technology Innovation Institute (TII) United Arab Emirates

Multi-Agent LLM Systems Swarm Robotics Multi-Agent Communication Memory-Augmented Agents

Open paper AskPDF arXiv

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.

Abstract

Explore similar work

Managing Context and Communication in Distributed Agentic UAV Swarms

LAEI: Layered Autonomous Edge Intelligence Framework for Robust UAV Swarm Operations

AeroWeaver: An Embodied-Agent Harness for Weaving Aerial Skills into Distributed, Adaptive Swarm Execution