SwarmChat Communication Systems

• SwarmChat is a family of decentralized communication systems designed for robotic swarms and P2P networks with context-aware, multimodal interaction.
• It spans diverse technical contexts including LLM-based human-swarm interfaces, anonymous mesh messaging via BLE, and motion-based signaling protocols.
• Its architecture integrates modules for context generation, intent recognition, task planning, and real-time adaptive feedback to optimize command execution and network scalability.

Searching arXiv for papers and related work on SwarmChat and adjacent swarm/mesh communication systems. SwarmChat denotes a family of swarm-oriented communication systems rather than a single canonical protocol. In the cited literature, the name is used most explicitly for an LLM-based, context-aware multimodal interaction system for robotic swarms, but it also appears as a derived design label for anonymous mesh messaging, trustless peer-to-peer chat, motion-signaled robot communication, and swarm-avatar telepresence. This suggests a unifying theme: communication is organized around decentralized collectives, whether the collective is a robotic swarm, a BLE mesh, a P2P overlay, or a formation of drones (Eumi et al., 21 Sep 2025).

1. Terminological scope and research contexts

The term spans several technical contexts with different substrates, threat models, and interaction modalities. The most direct usage is "SwarmChat: An LLM-Based, Context-Aware Multimodal Interaction System for Robotic Swarms" (Eumi et al., 21 Sep 2025). Other cited works recast adjacent systems as SwarmChat-style designs: anonymous mesh messaging over Bluetooth (Perry et al., 2022), three-overlay P2P social communication (Freitas, 2013), distributed secure P2P messaging (Halder et al., 2022), asynchronous communication in robot networks (Bouzid et al., 2010), explicit movement-based communication among stigmergic robots (0902.3549), and anthropomorphic swarm telepresence through gesture-recognizing drones (Baza et al., 2022).

Usage of "SwarmChat"	Core substrate	Source
LLM-based human-swarm interaction	Text, voice, teleop, ROS2, TurtleBots	(Eumi et al., 21 Sep 2025)
Anonymous mesh messaging design	BLE mesh, epidemic flooding, KEM, AEAD	(Perry et al., 2022)
P2P overlay communication design	Blockchain, DHT, BitTorrent-style swarms	(Freitas, 2013)
Trustless secure P2P messaging design	libp2p, Kademlia, WebRTC, DMQ	(Halder et al., 2022)
Motion-signaled robot messaging	RoboCast, stigmergy, asynchronous motion	(Bouzid et al., 2010, 0902.3549)
Embodied swarm telepresence interface	Body tracking, LSTM gesture recognition, drones	(Baza et al., 2022)

A related but distinct line of work is CHIRP and S-CHIRP, which introduce decentralized peer-to-peer communication for heterogeneous IoT devices with round-robin protection, enabling complete peer-to-peer communication of an $n$-agent network in as few as $n$ rounds, resilience to node loss and node re-entry, and theoretical scaling to swarms in the million-device range with memory constraints in the $< 10$ MB range. Only the abstract-level characterization is available in the cited record (Borowczak et al., 2017).

2. LLM-based robotic-swarm interaction system

In its most specific usage, SwarmChat is a three-layer pipeline that separates human interaction, LLM-based processing, and swarm execution. The user interaction layer exposes a Tkinter GUI for text entry, voice recording, and teleop keys, and displays keyword inputs $K$, candidate contexts $\{c_1,\dots,c_4\}$, similarity scores, suggested modalities, and real-time swarm status. The core layer contains Context Generator, Similarity Scoring, Intent Recognition, Task Planner, Modality Selector, Command Packaging, and Real-time Analytics and LLM Performance Logging. The robot swarm layer consists of ROS2 subscriber nodes on each TurtleBot, JSON parsing of $\{\text{target}, \text{command}, \text{modality}\}$, high-level to low-level conversion through Twist messages, and feedback publication on swarmchat/feedback (Eumi et al., 21 Sep 2025).

The processing chain is specified as

$$\text{User input (text | voice | teleop)} \rightarrow \text{extract keyword set } K \rightarrow f_{CG}(K)\to\{c_1,\dots,c_4\} \rightarrow \mathrm{score}_i=\mathrm{scale}\bigl(J(c_i,K)\bigr) \rightarrow \pi=f_{IR}(K) \rightarrow c^*=f_{TP}(c_i,\text{robot\_state}) \rightarrow m^*=\arg\max_m S(m\mid c^*,K),$$

followed by packaging of $\{\text{target}, c^*, m^*\}$ and publication via ROS2. The architecture therefore couples natural-language interpretation with runtime robot-state adaptation and modality recommendation, rather than treating command parsing as a purely static text classification problem (Eumi et al., 21 Sep 2025).

The abstract emphasizes four LLM-based modules—Context Generator, Intent Recognition, Task Planner, and Modality Selector—and situates them within a context-aware, multimodal interface intended to reduce cognitive load while preserving command flexibility. The stated interaction modalities are text, voice, and teleoperation, and the architecture offers both fixed and customizable command options (Eumi et al., 21 Sep 2025).

3. Core modules, scoring functions, and online adaptation

The Context Generator takes a keyword set $K=\{k_1,k_2,\dots\}$ and outputs four candidate contexts,

$f_{CG}(K)=\{c_1,c_2,c_3,c_4\}.$

Its implementation is prompt-driven: "Given keywords … generate four plausible natural-language commands." No embedding network or retrieval index is described. The system then applies Jaccard similarity between the keywords and each candidate,

$n$0

and rescales this to

$n$1

Module-specific bonus terms may then be added, with update rules

$n$2

This produces a lightweight adaptive mechanism rather than a trained end-to-end scorer (Eumi et al., 21 Sep 2025).

Intent Recognition is rule-based. If "patrol" is in $n$3, the system assigns Patrol mode; if "go" or "execute" is in $n$4, it assigns Navigation mode; otherwise it assigns General operation. The Task Planner refines the selected context $n$5 using real-time robot state $n$6, including position and battery, according to

$n$7

No explicit RL objective or heuristic formula beyond "incorporate real-time analytics" is detailed, and no custom classifier head or end-to-end fine-tuning is reported (Eumi et al., 21 Sep 2025).

The Modality Selector also uses hard rules. If $n$8, it suggests Teleop; if the context contains the verb "speak", it suggests Voice; otherwise it suggests Text. All modalities are normalized into a unified keyword set $n$9: raw GUI text, speech-to-text output, and teleop keyboard signals $< 10$0 are reduced to the same symbolic representation. User override remains possible, so the selector functions as an assistive recommender rather than an exclusive controller (Eumi et al., 21 Sep 2025).

Real-time adaptation is driven by feedback on the swarmchat/feedback topic. The LearningGraph logs execution success or failure and updates module weights using

$< 10$1

This places the system between a static rule engine and a fully learned policy: adaptation occurs online, but through stochastic gradient-like weight updates over modular scores rather than end-to-end optimization (Eumi et al., 21 Sep 2025).

4. Mesh, overlay, and trustless-network interpretations

One research line adapts SwarmChat to anonymous mesh messaging over Bluetooth. The threat model includes local eavesdroppers on Bluetooth Low Energy links, a fraction $< 10$2 of compromised peers, global passive traffic analysis, active jamming and message injection, and limited selective denial-of-service. Required guarantees include sender anonymity, recipient anonymity, and unlinkability against PPT adversaries. Broadcast proceeds through a push-pull epidemic model over an undirected BLE-mesh graph $< 10$3, with random delay $< 10$4, per-neighbor gossip probability $< 10$5, duplicate suppression through cached message identifiers, and TTL bounded by $< 10$6. The expected reached fraction is approximated by

$< 10$7

with fixed point $< 10$8. In the cited evaluation, $< 10$9, average degree $K$0, $K$1, TTL$K$2, and $K$3 ms; with $K$4 and $K$5, delivery latency is approximately $K$6 s and bandwidth overhead approximately $K$7. Group messaging is formulated either through subset-cover broadcast encryption or pairwise group key agreement (Perry et al., 2022).

A second line derives SwarmChat from the three-overlay architecture of a P2P microblogging platform. Here the system is decomposed into a Bitcoin-inspired registration and authentication overlay, a Kademlia-style DHT for key/value storage and tracker lookup, and BitTorrent-style swarms for near-instant notification delivery. Identity binding uses

$K$8

registration records

$K$9

and a lightweight proof-of-work condition

$\{c_1,\dots,c_4\}$0

For anti-abuse control, the rate limit on new users’ "have" announcements is

$\{c_1,\dots,c_4\}$1

DHT routing cost is $\{c_1,\dots,c_4\}$2, with expected lookup latency

$\{c_1,\dots,c_4\}$3

Swarm propagation is estimated by

$\{c_1,\dots,c_4\}$4

This interpretation treats chat rooms as swarm overlays and separates registration, storage, and dissemination into distinct layers (Freitas, 2013).

A third line treats SwarmChat as a trustless, fully peer-to-peer secure messaging system built on a flat mesh with Kademlia-style discovery, content addressing, WebRTC direct routing, pinning, and a Distributed Message Queue. Each message or file is hashed with SHA-256 to obtain a 256-bit Content ID, and offline delivery is mediated through a DHT-backed index keyed by the recipient identity. The routing distance is the XOR metric,

$\{c_1,\dots,c_4\}$5

and expected hop count is

$\{c_1,\dots,c_4\}$6

Direct WebRTC DataChannel delivery is the low-latency fast path, while DHT lookup is the fallback path with

$\{c_1,\dots,c_4\}$7

End-to-end encryption uses long-term curve25519 ECDH public keys, ephemeral-static key agreement $\{c_1,\dots,c_4\}$8, HKDF-based shared-secret derivation, and AEAD encryption. In the cited prototype, 500 messages of 50–500 bytes yield average latency of $\{c_1,\dots,c_4\}$9 s/msg on the WebRTC DataChannel, compared with $\{\text{target}, \text{command}, \text{modality}\}$0 s/msg for WhatsApp and $\{\text{target}, \text{command}, \text{modality}\}$1 s/msg for the IPFS fallback channel (Halder et al., 2022).

5. Motion-based swarm communication and embodied telepresence

In robot-network research, SwarmChat can also denote communication implemented through movement rather than conventional packets. The RoboCast abstraction allows non-oblivious, anonymous robots with visibility sensors and no common coordinate system to asynchronously exchange information. Each robot invokes $\{\text{target}, \text{command}, \text{modality}\}$2 and peers eventually execute $\{\text{target}, \text{command}, \text{modality}\}$3. The abstraction requires validity and termination, and it is used to broadcast local coordinate systems, implement deterministic asynchronous gathering, and exchange binary information. Collision avoidance is enforced by the local zone of motion

$\{\text{target}, \text{command}, \text{modality}\}$4

where

$\{\text{target}, \text{command}, \text{modality}\}$5

This yields collision-free communication-compatible motion while preserving the logic of directional reversals and line-based broadcasts (Bouzid et al., 2010).

An earlier stigmergic communication line formalizes explicit one-to-one communication among deaf and dumb mobile robots through movement-signals. It covers synchronous protocols with visible IDs, anonymous synchronous protocols under common direction and chirality or chirality only, and asynchronous protocols with implicit acknowledgments. In the ID-based synchronous protocol, each robot slices its local granular into $\{\text{target}, \text{command}, \text{modality}\}$6 equal angular sectors and encodes a bit $\{\text{target}, \text{command}, \text{modality}\}$7 for target $\{\text{target}, \text{command}, \text{modality}\}$8 through a movement vector

$\{\text{target}, \text{command}, \text{modality}\}$9

In the asynchronous setting, the key idea is that observing the peer move twice acts as an implicit acknowledgment. The general $$\text{User input (text | voice | teleop)} \rightarrow \text{extract keyword set } K \rightarrow f_{CG}(K)\to\{c_1,\dots,c_4\} \rightarrow \mathrm{score}_i=\mathrm{scale}\bigl(J(c_i,K)\bigr) \rightarrow \pi=f_{IR}(K) \rightarrow c^*=f_{TP}(c_i,\text{robot\_state}) \rightarrow m^*=\arg\max_m S(m\mid c^*,K),$$0-robot protocol combines this acknowledgment mechanism with sector-based addressing, yielding worst-case $$\text{User input (text | voice | teleop)} \rightarrow \text{extract keyword set } K \rightarrow f_{CG}(K)\to\{c_1,\dots,c_4\} \rightarrow \mathrm{score}_i=\mathrm{scale}\bigl(J(c_i,K)\bigr) \rightarrow \pi=f_{IR}(K) \rightarrow c^*=f_{TP}(c_i,\text{robot\_state}) \rightarrow m^*=\arg\max_m S(m\mid c^*,K),$$1 steps per bit. In this usage, SwarmChat is literally a communication substrate built from locomotion, observability, and geometric coding (0902.3549).

A further extension is affective telepresence through an anthropomorphic swarm avatar. SwarMan uses nine Crazyflie nano-quadcopters, VICON motion capture, MediaPipe Holistic body tracking, and an LSTM-based gesture recognizer trained on 600 short videos with five emotional gestures. The input is a sequence of 30 frames, each with 27 features, and the network consists of an LSTM layer with 128 hidden units and return_sequences=True, dropout with $$\text{User input (text | voice | teleop)} \rightarrow \text{extract keyword set } K \rightarrow f_{CG}(K)\to\{c_1,\dots,c_4\} \rightarrow \mathrm{score}_i=\mathrm{scale}\bigl(J(c_i,K)\bigr) \rightarrow \pi=f_{IR}(K) \rightarrow c^*=f_{TP}(c_i,\text{robot\_state}) \rightarrow m^*=\arg\max_m S(m\mid c^*,K),$$2, a second LSTM layer with 64 hidden units, and a dense softmax output over $$\text{User input (text | voice | teleop)} \rightarrow \text{extract keyword set } K \rightarrow f_{CG}(K)\to\{c_1,\dots,c_4\} \rightarrow \mathrm{score}_i=\mathrm{scale}\bigl(J(c_i,K)\bigr) \rightarrow \pi=f_{IR}(K) \rightarrow c^*=f_{TP}(c_i,\text{robot\_state}) \rightarrow m^*=\arg\max_m S(m\mid c^*,K),$$3 classes. The loss is categorical cross-entropy,

$$\text{User input (text | voice | teleop)} \rightarrow \text{extract keyword set } K \rightarrow f_{CG}(K)\to\{c_1,\dots,c_4\} \rightarrow \mathrm{score}_i=\mathrm{scale}\bigl(J(c_i,K)\bigr) \rightarrow \pi=f_{IR}(K) \rightarrow c^*=f_{TP}(c_i,\text{robot\_state}) \rightarrow m^*=\arg\max_m S(m\mid c^*,K),$$4

optimized with Adam at learning rate $$\text{User input (text | voice | teleop)} \rightarrow \text{extract keyword set } K \rightarrow f_{CG}(K)\to\{c_1,\dots,c_4\} \rightarrow \mathrm{score}_i=\mathrm{scale}\bigl(J(c_i,K)\bigr) \rightarrow \pi=f_{IR}(K) \rightarrow c^*=f_{TP}(c_i,\text{robot\_state}) \rightarrow m^*=\arg\max_m S(m\mid c^*,K),$$5. The reported test accuracy is $$\text{User input (text | voice | teleop)} \rightarrow \text{extract keyword set } K \rightarrow f_{CG}(K)\to\{c_1,\dots,c_4\} \rightarrow \mathrm{score}_i=\mathrm{scale}\bigl(J(c_i,K)\bigr) \rightarrow \pi=f_{IR}(K) \rightarrow c^*=f_{TP}(c_i,\text{robot\_state}) \rightarrow m^*=\arg\max_m S(m\mid c^*,K),$$6, and real-time user-study accuracy is $$\text{User input (text | voice | teleop)} \rightarrow \text{extract keyword set } K \rightarrow f_{CG}(K)\to\{c_1,\dots,c_4\} \rightarrow \mathrm{score}_i=\mathrm{scale}\bigl(J(c_i,K)\bigr) \rightarrow \pi=f_{IR}(K) \rightarrow c^*=f_{TP}(c_i,\text{robot\_state}) \rightarrow m^*=\arg\max_m S(m\mid c^*,K),$$7. Emotion classes are mapped to LED colors—Happy to green, Sad to blue, Angry to red, Neutral to white, and Confused to yellow—while collision avoidance uses artificial potential fields with

$$\text{User input (text | voice | teleop)} \rightarrow \text{extract keyword set } K \rightarrow f_{CG}(K)\to\{c_1,\dots,c_4\} \rightarrow \mathrm{score}_i=\mathrm{scale}\bigl(J(c_i,K)\bigr) \rightarrow \pi=f_{IR}(K) \rightarrow c^*=f_{TP}(c_i,\text{robot\_state}) \rightarrow m^*=\arg\max_m S(m\mid c^*,K),$$8

This makes SwarmChat an embodied, multimodal communication channel in which motion, formation, and illumination jointly encode affect (Baza et al., 2022).

6. Evaluation, limitations, and interpretive boundaries

The robotic-swarm SwarmChat system reports a preliminary evaluation in which four sample keyword sets were run through each module in real time across three modalities and three TurtleBots. The metrics were Jaccard-based similarity score, module-wise suggestion versus user choice, and satisfaction rating on a four-point scale. Perfect scores of $$\text{User input (text | voice | teleop)} \rightarrow \text{extract keyword set } K \rightarrow f_{CG}(K)\to\{c_1,\dots,c_4\} \rightarrow \mathrm{score}_i=\mathrm{scale}\bigl(J(c_i,K)\bigr) \rightarrow \pi=f_{IR}(K) \rightarrow c^*=f_{TP}(c_i,\text{robot\_state}) \rightarrow m^*=\arg\max_m S(m\mid c^*,K),$$9 yielded Teleop suggestions and 100% user acceptance in patrol commands, while lower similarity scores of $\{\text{target}, c^*, m^*\}$0 led to varied modality choices and medium satisfaction. The GUI presents four candidate contexts with similarity scores, suggested modality with override buttons, and controls to inspect published versus received commands. Satisfaction levels are reported as Very High, High, Medium, and Low, and no formal NASA-TLX scores have yet been collected (Eumi et al., 21 Sep 2025).

The same source states several limitations directly. Intent recognition is rule-based; no classifier head or loss function is reported for that module; no end-to-end fine-tuning is described; and the task planner has no explicit RL objective. Future work includes learning-based intent recognition with trainable heads and cross-entropy loss, larger-scale user studies with ANOVA, $\{\text{target}, c^*, m^*\}$1-tests, and $\{\text{target}, c^*, m^*\}$2-values, collection of standard cognitive-load metrics such as NASA-TLX, and integration of additional modalities including eye-tracking and BCI. A plausible implication is that the present system should be read as a modular prototype rather than a closed, benchmarked HSI stack (Eumi et al., 21 Sep 2025).

A common misconception is to treat SwarmChat as a standardized protocol with a single architecture. The cited record does not support that interpretation. Instead, it supports several technically distinct usages: LLM-mediated command and control for robotic swarms, anonymity-preserving mesh messaging, trustless P2P chat over overlay networks, explicit communication through robot motion, and swarm-avatar telepresence. This suggests that "SwarmChat" functions primarily as a descriptive label for communication systems whose semantics and robustness emerge from swarm organization, not as a fixed protocol suite.

Another boundary concerns secure swarm communication in adjacent literature. CHIRP and S-CHIRP provide an abstract-level example of decentralized, resilient, complete peer-to-peer communication for heterogeneous IoT swarms in as few as $\{\text{target}, c^*, m^*\}$3 rounds, including resilience to node loss and node re-entry and theoretical scaling to the million-device range with memory constraints below 10 MB. However, the detailed formulas, cryptographic design, churn handling, and performance evaluation are not available in the cited record. Accordingly, they should be regarded as related background rather than a fully exposable SwarmChat instance (Borowczak et al., 2017).