Optimizing Data Distribution Service Discovery for Swarm Unmanned Aerial Vehicles Through Preloading and Network Awareness

Highlights

What are the main findings?

• A network-aware DDS mechanism significantly reduces traffic congestion during UAV swarm discovery.
• EDP preloading minimizes initialization overhead by embedding known endpoint data before deployment.

What is the implication of the main finding?

• The proposed method enables more stable and scalable communication in dense UAV swarms.
• This approach facilitates real-time coordination without requiring changes to existing DDS architecture.

Abstract

Collaborative unmanned aerial vehicle (UAV) swarm operations using the open-source PX4–ROS2 system have been extensively studied for reconnaissance and autonomous missions. PX4–ROS2 utilizes data distribution service (DDS) middleware to ensure network flexibility and support scalable operations. DDS enables decentralized information exchange through its discovery protocol. However, in dense swarm environments, the default initialization process of this protocol generates considerable communication overhead, which hinders reliable peer detection among UAVs. This study introduces an optimized DDS discovery scheme incorporating two key strategies: a preloading method that embeds known participant data before deployment, and a dynamic network awareness approach that regulates discovery behavior based on real-time connectivity. Integrated into PX4–ROS2, the proposed scheme was assessed through both simulations and real-world testing. Results demonstrate that the optimized discovery process reduced peak packet traffic by over 90% during the initial exchange phase, thereby facilitating more stable and scalable swarm operations in wireless environments.

1. Introduction

Recent advances in swarm operation technologies have driven active research into unmanned aerial vehicle (UAV) swarms across a range of domains, including reconnaissance, disaster response, logistics, and agriculture [1]. UAV swarms are increasingly utilized in environments that are difficult or hazardous for human access, such as large-scale farmland for precision pesticide spraying [2], automated inventory management in warehouses [3,4], and rapid search-and-rescue operations in military or disaster zones [5]. These applications have garnered significant interest, as UAVs offer safe and efficient task execution while improving automation and responsiveness in high-risk settings.

In swarm systems with simultaneous UAV operations, effective communication and information sharing require a flexible framework to maintain both stability and scalability. As a result, substantial research has examined the data distribution service (DDS) protocol, developed by the Object Management Group (OMG), as a mechanism for dynamic data exchange in distributed UAV networks [6]. DDS supports data-centric communication, enabling UAVs to dynamically join or exit the swarm while autonomously managing message formats and ensuring connection reliability. It also adopts a publish-subscribe model, which decouples data producers from consumers and supports fault-tolerant architectures with minimal risk of single points of failure (SPOF) [7].

DDS facilitates distributed information exchange among UAVs through its discovery mechanism [8,9,10], which automatically detects participating UAVs and exchanges quality of service (QoS) settings and topic metadata. However, in swarm operations, the default discovery process often generates excessive traffic during initialization, often resulting in failures to recognize certain UAVs. As the swarm size grows, network traffic increases exponentially, which can disrupt the discovery process, particularly in wireless environments.

Although several previous studies have validated the performance of DDS in wired networks, most have not addressed the dynamic mobility and variable traffic patterns characteristic of real-world wireless deployments [11,12,13,14,15,16,17,18]. This gap underscores the need for optimized DDS protocols specifically tailored to UAV swarms operating under fluctuating wireless conditions.

To address this limitation, the present study introduces a DDS-based communication optimization approach for UAV swarms in wireless environments. First, we propose a network awareness strategy that monitors the received signal strength indicator (RSSI) in real time and dynamically allocates traffic based on link quality to alleviate traffic bursts. Second, we define a preloading discovery that embeds essential endpoint data into UAV firmware when the swarm topology is predefined, thereby reducing initial communication overhead.

We implemented the proposed methods on the PX4 (v1.14)–ROS2 (Humble) platform [19,20,21] and validated their effectiveness through both simulations and real-world testing. The results demonstrated that the methods significantly reduced the initial communication overhead and enhanced transmission reliability, even in dense and dynamic UAV swarm conditions.

The main contributions of this study are as follows:

• We propose a network-awareness method in which each UAV monitors RSSI values to assess network conditions in real time and adaptively allocate traffic, thereby mitigating instability in wireless environments.
• We propose a preloading discovery method that embeds essential discovery metadata into each UAV ahead of deployment when the swarm configuration is predefined, simplifying communication procedures and reducing initial traffic overhead.
• We implement the proposed methods on the PX4–ROS2 platform, a widely adopted framework in unmanned systems, and evaluate their performance through both simulations and real-world tests. The results show that the combined approach significantly reduces initial packet load and enhances the success rate of wireless communication, supporting scalable swarm operation.

The remainder of this paper is organized as follows: Section 2 reviews previous research on DDS communication in unmanned systems. Section 3 introduces the PX4–ROS2 platform and relevant technical background. Section 4 details the proposed network-awareness algorithm and preloading discovery scheme. Section 5 presents the methodology and results of simulation and real-world experiments. Section 6 concludes the study and discusses directions for future research.

2. Related Works

In UAV swarm operations, DDS has emerged as a core middleware due to its ability to support real-time data exchange and interoperability among heterogeneous UAVs. Enabled by a diverse set of QoS policies defined by the OMG, DDS can flexibly adapt to dynamic network environments. Maruyama et al. compared multiple DDS implementations under different configurations, analyzing how QoS parameters influence performance [22]. Kronauer et al. evaluated latency and throughput across DDS vendors in multi-UAV scenarios, confirming that system performance is highly sensitive to both QoS settings and underlying network conditions [12]. García-Valls et al. integrated DDS with partitioned distributed systems to improve throughput via data segregation [13], while Yoon et al. proposed a dynamic QoS selection method based on runtime monitoring data [14]. However, these studies primarily focused on conceptual designs and lacked validation in practical UAV deployments.

Some efforts have extended DDS functionality. Auliya et al. used subscriber feedback to dynamically adjust publisher transmission rates, addressing the trade-off between throughput and reliability in lossy wireless environments [15]. Martin-Carrascosa et al. developed a congestion detection algorithm that throttles publishers to optimize bandwidth, though their method only operates after unicast links are established, limiting its applicability in dynamic UAV swarms.

Despite these developments, many DDS optimization strategies rely on post-connection monitoring or QoS parameter tuning and often neglect the surge of network-awareness traffic during initial swarm formation, particularly in wireless contexts. Almadani et al. experimentally evaluated DDS reliability over wireless links and observed performance variations tied to QoS settings [17]. Boonma and Suzuki proposed TinyDDS, a lightweight publish-subscribe middleware for resource-constrained sensor networks, highlighting DDS’s feasibility in limited environments [18]. However, structural solutions to reduce network-awareness traffic at initialization remain lacking.

Several studies have aimed to reduce the overhead associated with the DDS discovery process. Sánchez-Monedero et al. introduced a Bloom filter-based protocol to compress endpoint discovery protocol (EDP) traffic [23], while Nwadiugwu et al. enhanced this approach using threshold-based Bloom filters to reduce false positives and memory usage [24]. However, these techniques have been validated primarily in wired network settings, and their performance under congested wireless conditions remains uncertain.

Choi et al. integrated software-defined networking (SDN) with DDS to prioritize real-time flows at the switch level and improve delivery reliability [25]. Other studies have proposed RSSI-based adaptations for mobile platforms, often targeting transmission intervals or data-plane behavior. However, these approaches do not directly address the network-awareness phase of DDS initialization.

In addition, Awada et al. proposed a collaborative scheduling framework called EdgeDrones, which jointly optimizes UAV task execution and data transmission to maximize resource utilization in multi-location aerial missions [26]. This approach emphasizes high-level orchestration and flight planning across edge computing environments rather than addressing the low-level network congestion that occurs immediately prior to communication. In contrast, the present study focuses on the discovery stage—the initial control-plane burst triggered when UAVs join a domain. As such, the two approaches operate at different layers of the system stack and can serve complementary roles.

To address these limitations, the present study adopts DDS as the core communication middleware and implements the proposed network-awareness and preloading discovery methods on a PX4–ROS 2-based UAV platform [27]. The system integrates a DDS for extremely Resource-Constrained Environments (XRCE-DDS) bridge to connect PX4 with the DDS infrastructure, enabling lightweight UAVs to participate in swarm communication. This work provides practical evidence and actionable guidelines for deploying DDS in large-scale UAV networks over wireless links, validated through both simulation and real-world testing.

Table 1. Comparison of existing DDS strategies and the proposed method.

Category [Refs.]	Pros	Cons	Proposed
QoS tuning [12,13,14,22]	Low learning cost and straightforward parameter configuration	Limited to predefined QoS fields and static after deployment	Network awareness–based structural tuning enables fine-grained adaptation during runtime
Protocol redesign [15,16]	Sliding-window or rate control methods improve throughput	Takes effect only after discovery and applies only to unicast phase	Controls traffic dynamically from the very first discovery packet
Wireless context [17,18]	Real Wi-Fi and WSN testbeds quantify the QoS impact under wireless conditions	Lacks structural solutions to suppress burst traffic during discovery	Introduces a burst-reduction strategy tailored for UAV mesh using preloading and network awareness
Discovery overhead [23,24]	Bloom filters or threshold-based filtering reduce discovery message count	Evaluated only on wired local area network; impact on wireless networks remains unverified	Suppresses discovery bursts in wireless UAV swarms through a preloading and awareness-guided scheme
Network-flow control [25]	SDN prioritizes DDS traffic flows, reducing latency under congestion	Requires dedicated switches and lacks awareness prior to discovery initiation	Achieves comparable latency benefits without SDN infrastructure by incorporating network awareness

presents a structured comparison between existing DDS strategies and the proposed method, highlighting how the network awareness and EDP preloading enable scalable and efficient swarm initialization in wireless UAV environments.

3. Preliminaries

This section introduces the experimental platform by presenting the PX4–ROS2 system architecture and outlining the DDS discovery process to characterize the system’s communication behavior.

3.1. PX4–ROS2

As shown in Figure 1, PX4–ROS2 is an open-source UAV system built on the PX4 flight controller and ROS 2 middleware, making it highly suitable for swarm UAV operations. PX4 is a widely adopted open-source flight control system used across a range of UAV types and other unmanned platforms, including unmanned ground vehicles (UGVs) and unmanned surface vehicles (USVs) [28]. Developed in C++, PX4 features a modular architecture comprising a hardware abstraction layer (HAL), sensor drivers, flight control firmware, and communication modules. These components interact via uORB, a lightweight publish-subscribe messaging framework optimized for real-time performance with minimal overhead  [28].

ROS 2 is a middleware framework redesigned to support industrial and distributed robotic systems, addressing several limitations of ROS 1 [27]. It uses a publish-subscribe architecture based on DDS, which significantly enhances real-time performance, scalability, and system-level security. Within the PX4–ROS2 integration, DDS enables external communication through the real-time publish-subscribe (RTPS) protocol, ensuring high reliability for control tasks and fault tolerance. This architecture enables seamless message exchange, leveraging the modular and publish-subscribe designs common to both PX4 and ROS 2.

The XRCE-DDS protocol enables efficient, lightweight communication between PX4 and ROS 2. Designed to support DDS communication under severe resource constraints [29], XRCE-DDS consists of an XRCE Client operating within the PX4 environment and an XRCE Agent running on the ROS 2 side. The client issues topic-related requests, while the agent mediates communication with the broader DDS domain. By offloading endpoint management to the agent, the client minimizes resource consumption, allowing even highly constrained devices to participate in DDS-based data exchange  [30]. As shown in Figure 1, this architecture forms a client–agent bridge that enables robust, scalable communication for swarm UAV operations.

3.2. DDS (Data Distribution Service)

The OMG-standardized DDS operates on the data-centric publish-subscribe (DCPS) model [31], providing flexible and efficient communication via a publish-subscribe architecture. DDS middleware supports loose coupling by enabling data exchange through topic names without requiring explicit endpoint configuration. Internally, DDS uses the RTPS protocol as its communication foundation [32]. RTPS is specifically designed for real-time systems, offering low latency and high reliability. It also provides fine-grained control over transmission behavior via QoS parameters, including reliability, priority, bandwidth, and latency. As a result, DDS is widely adopted in systems requiring both real-time responsiveness and communication reliability.

To enable decentralized information sharing among nodes, DDS initiates a discovery process. During this phase, nodes automatically detect one another and exchange essential QoS settings and topic metadata. As shown in Figure 2, the discovery process consists of the participant discovery protocol (PDP) and the EDP [32]. PDP communicates each node’s metadata—including domain ID, protocol version, vendor ID, QoS policies, and locators—using multicast or predefined unicast messages. Each node simultaneously identifies and announces its presence to other participants in the same domain. EDP then examines publication and subscription endpoints, verifying compatibility between DataWriters and DataReaders based on their data types and QoS configurations, thereby completing the communication link setup.

Because DDS does not rely on a centralized controller or any SPOF, all nodes participate equally in both PDP and EDP exchanges. This decentralized architecture enables self-configuration and supports flexible, scalable communication, forming the foundation for efficient data exchange across nodes in the DDS network [7,33].

As Figure 3 illustrates, DDS supports three primary discovery architectures [34]. The Simple method involves each node engaging in both PDP and EDP communication with every other node. This approach requires no additional configuration and automatically adapts to changes in endpoints. However, as the number of nodes increases, multicast and broadcast traffic grow exponentially, leading to congestion and collisions in wireless environments.

In the Server discovery method, a designated server node manages PDP communication. All other nodes interact solely with this server, which collects and redistributes participant information. This approach reduces overall PDP traffic and simplifies participant registration and deregistration. However, it introduces an SPOF since a failure at the server can disrupt the entire system.

The Static discovery method uses pre-defined EDP tables on each node, minimizing PDP exchange during runtime. By eliminating dynamic EDP, this method generates the least traffic, thereby reducing both latency and network congestion [35]. However, it requires manual updates whenever shared endpoint information changes, making it unsuitable for environments that demand frequent reconfiguration.

Default DDS discovery mechanisms scale poorly in wireless networks, as traffic volume increases exponentially with node count. These limitations are particularly problematic in swarm operations, where communication conditions change dynamically throughout missions. To address this, the present study proposes a new discovery method that incorporates network awareness to detect changing conditions and minimize unnecessary communication within the swarm.

4. Proposed Discovery Mechanism

To enable collaborative mission execution among multiple UAVs, information must be exchanged over a wireless network. However, in swarm environments where many UAVs operate concurrently, such communication is vulnerable to packet loss and limited bandwidth. While decentralized architectures are generally better suited for stable mission execution than centralized systems, they inevitably result in increased communication volume, which must be controlled to ensure effective UAV swarm operations.

This study introduces an optimized DDS discovery mechanism specifically designed for UAV swarm environments. The proposed method integrates network awareness and EDP preloading discovery to minimize signaling overhead. It dynamically regulates discovery traffic based on real-time network conditions and preloads known peer information when the swarm topology is predetermined.

In the standard Simple discovery protocol provided by DDS (Figure 4a), each UAV attempts to connect with all visible participants in the domain during the PDP phase. As shown in Section 5, this results in a burst of traffic during initial discovery—especially when using FastDDS. Such communication surges are problematic in UAV networks, where bandwidth is limited. The increased packet loss during this phase may prevent some UAVs from being discovered, thereby hindering collaboration and system synchronization.

4.1. Network Awareness

To address this challenge, we introduce a mechanism that adjusts discovery timing based on real-time network conditions. Prior to joining the domain, each UAV is assigned a Passive Timer derived from its RSSI level, as illustrated in Figure 4b. The Passive Timer calculation depends on the wireless driver’s capability. If link quality ($LQ$) is available, it is normalized by the maximum link quality ($L{Q}_{max}$) as follows:

$$Q_i={\displaystyle \frac{LQ}{L{Q}_{max}}},0\le Q_i\le 1$$

(1)

If only RSSI and noise measurements are available, the signal-to-noise ratio (SNR) is computed and normalized within the 25–40 dB range. This range is chosen because communication quality is generally stable above 25 dB, while link improvements tend to saturate beyond 40 dB [36,37,38]:

$$Q_i=clamp\left({\displaystyle \frac{SNR}{40}},0,1\right)$$

(2)

Based on $Q_i$, the Passive Timer $T_{\mathrm{passive}}^{\left(i\right)}$ is sampled from a uniform distribution within a predefined bounded interval:

$$T_{passive}^{\left(i\right)}\sim \mathcal{U}\left[0,T_{min}+(T_{max}-T_{min})(1-Q_i)\right](T_{min}=8\mathrm{s},T_{max}=20\mathrm{s})$$

(3)

This timer delays PDP packet transmission after a UAV joins the domain, thereby suppressing initial traffic bursts. During this delay period, the UAV remains silent but listens for incoming PDP traffic to passively identify nearby UAVs. Once these UAVs are detected, they establish partial unicast connections with them. Because the Passive Timer is applied only once at the DDS control plane during system initialization, it does not affect latency requirements on the user plane.

As shown in Figure 4b, unlike the standard DDS domain, where multicast discovery targets all UAVs simultaneously, the proposed scheme defers connection attempts to unresponsive UAVs until the Passive Timer expires. This staggered approach mitigates congestion and reduces the likelihood of packet collisions during the critical startup phase.

Building on the OMG DDS reference architecture Figure 5a [31,32], the proposed DDS discovery mechanism with network awareness introduces a cross-layer design (Figure 5b) to enhance congestion control and scalability in wireless environments. While the core DDS layered structure remains intact, an integrated Network Monitor module within the DDS domain layer continuously reads RF receiver status (e.g., RSSI) from the physical layer, providing real-time LQ feedback to the RTPS layer. Based on this input, a Passive Filter Module embedded in RTPS dynamically adjusts the timing of initial PDP advertisements, drawing inspiration from adaptive-beaconing strategies in VANETs [39].

The Passive Filter modulates the broadcast interval of PDP messages generated by the RTPS Writer. After joining the domain, a UAV initially remains silent and refrains from broadcasting its presence. Instead, it passively listens for PDP packets transmitted by neighboring UAVs to identify potential peers. Upon detection, the UAV establishes partial connections via unicast. The standard PDP procedure is then triggered only after the Passive Timer expires, specifically targeting UAVs that have not yet been discovered.

This staggered discovery process reduces initial traffic surges and mitigates packet collisions during the critical startup phase. By deferring explicit presence advertisements, UAVs can operate within the domain while maintaining minimal overhead. This approach supports scalable and stable swarm initialization, even in congested wireless environments. The duration of the Passive Timer does not impact the latency budgets guaranteed by underlying radio-layer protocols. Further explanation is provided in Appendix A.

4.2. EDP Preloading Discovery

A typical DDS discovery mechanism sequentially executes both the PDP and the EDP to identify all participants and their communication endpoints within a domain. As shown in Figure 6a, the standard Simple discovery protocol requires each UAV to broadcast its PDP packet upon joining the domain and then exchange endpoint metadata with all other participants via EDP. As the number of UAVs increases, EDP traffic grows exponentially. In wireless swarm environments, this burst of control traffic during the discovery phase can lead to severe network congestion and degraded system performance.

To address this challenge, we propose a preloading discovery scheme (Figure 6b) that leverages prior knowledge of topic-level communication patterns defined during mission planning. Required EDP metadata are embedded into each UAV before deployment, enabling only the PDP phase to be executed at runtime for presence announcement.

The proposed method inherits the structural foundation of Static discovery by provisioning endpoint metadata before node activation [35,40], but it differs significantly in how EDP data are managed. In Static discovery, each UAV reads from a predefined EDP table and establishes endpoint connections sequentially. As the number of UAVs and table entries increases, communication overhead scales proportionally, generating substantial discovery traffic.

In contrast, the preloading discovery method batches all EDP entries before runtime and initiates communication only after the metadata are embedded. This eliminates the need for sequential endpoint negotiation and its associated traffic. Once the PDP packet announces a UAV’s presence, data exchange begins immediately using the preloaded EDP information, without incurring further traffic for endpoint matching. Consequently, preloading discovery significantly reduces overall discovery traffic compared with Static discovery and supports scalable, reliable swarm initialization in dense wireless environments where real-time responsiveness is essential.

As depicted in Figure 7, the proposed preloading scheme bypasses the EDP phase entirely. This structural advantage prevents message bursts during swarm initialization and ensures that discovery traffic does not scale exponentially with the number of UAVs.

Figure 8 presents the measured discovery-phase traffic under different DDS discovery schemes. While Static Discovery still incurs additional PDP messages due to incremental EDP table updates, the preloading approach synchronizes all metadata once before runtime, resulting in the lowest packet volume among the tested methods. This confirms that preloading discovery is highly effective in mitigating traffic and supporting scalable swarm deployment in bandwidth-constrained networks.

5. Experiments

To validate the proposed DDS discovery mechanism, we conducted simulation-based and real-world experiments. The simulation study implemented a simplified ROS 2-based environment to evaluate the discovery algorithm. Each UAV operated under identical wireless network conditions, modeled using the NS-3 simulator [41]. Additionally, the PX4–ROS 2 UAV system was integrated with the Gazebo simulator [42] to emulate realistic swarm operations. UAVs were arranged linearly so that packet loss varied with distance, as reflected in the NS-3 wireless settings. For real-world validation, we deployed ten PX4–ROS 2 UAVs in a live Wi-Fi environment. For comprehensive evaluation, the proposed discovery methods were compared with widely used DDS implementations in ROS 2: eProsima Fast DDS, Eclipse Cyclone DDS, and RTI Connext DDS. These vendors are officially designated as Tier 1 middleware in ROS 2, offering full compatibility and long-term support [43]. Among them, Fast DDS serves as the default middleware in ROS 2 and was selected as the baseline for our implementation. Cyclone DDS was adopted as the default in the ROS 2 Galactic distribution, while Connext DDS is the first commercial DDS implementation and has been supported since the early ROS 2 releases. Although not a standard ROS 2 DDS implementation, Zenoh was also included as a reference due to its built-in discovery optimization features and increasing attention as a lightweight alternative. Other DDS vendors, such as OpenSplice or GurumDDS, were excluded from evaluation due to discontinued support or limited adoption in ROS 2 environments.

5.1. Simple Simulation Experiments

A simplified simulation environment, illustrated in Figure 9, was developed to evaluate the discovery and data exchange performance of DDS communication in ROS 2. The experiment was structured so that a single subscriber received messages from multiple publishers. NS-3 was configured to emulate an IEEE 802.11 wireless network representative of typical UAV mission environments. Each ROS 2 UAV node was containerized using Docker to ensure isolation from external interference. A Tap Bridge interface connected the containers to the NS-3 network, acting as a bridge between the ROS 2 processes and the simulated wireless environment. This setup enabled realistic testing of discovery behavior and message delivery under wireless conditions. Each publisher transmitted a simple string message on a shared topic. The experimental procedure was as follows:

• Each ROS 2 Publisher transmitted one message per second over the NS-3 wireless channel.
• The Subscriber recorded all received messages and monitored internal metrics such as packet loss and transmission attempts using NS-3.

All simulations were conducted on an Ubuntu 22.04 Linux system with an AMD Ryzen 9 7900 CPU (12 cores, 24 threads) and 32 GB of RAM. The number of UAVs varied from 2 to 30, and traffic characteristics were analyzed during the initial discovery phase. Each DDS implementation was tested with its default QoS configuration. A test was considered successful if the subscriber received the 20th message from each publisher within the defined transmission window.

To evaluate the proposed discovery mechanism, we independently tested the network awareness and preloading methods, as well as their combined implementation. The experiment involved 20 UAVs configured with Best Effort QoS, which does not guarantee packet delivery. As shown in Figure 10a, the standard discovery approach produced a traffic spike of up to 70,000 packets per second during the initial phase. In contrast, the network awareness method regulated message transmission based on real-time connectivity, reducing peak traffic to approximately 20,000 packets per second. This reduction significantly mitigated packet loss under constrained wireless bandwidth.

The preloading method reduced the total packet count by embedding EDP data into UAVs prior to deployment. However, as shown in Figure 10b, even with preloading, the packet ratio sharply increased with the number of UAVs, indicating that scalability remained a concern. The combined network awareness and preloading approach yielded the most favorable results, minimizing both the peak packet rate and total traffic ratio. These findings demonstrate superior communication efficiency in dense swarm environments.

We evaluated the discovery success rate, defined as the proportion of test runs in which all UAVs in the swarm successfully recognized one another during the discovery phase. Each configuration was tested 20 times, and the average number of successful runs was reported as the success rate. Under Best Effort QoS, as shown in Figure 11a, Fast DDS exhibited a sharp decline in success rate when the number of UAVs exceeded five. Connext DDS and Cyclone DDS experienced similar drops beyond 10 and 13 UAVs, respectively. In contrast, the proposed discovery mechanism maintained stable peer recognition up to 20 UAVs and achieved over 80% success even at 30 UAVs, outperforming all other DDS implementations. Comparable results were observed under Reliable QoS settings, as illustrated in Figure 11b. These findings demonstrate that the proposed method enhances discovery stability by reducing retransmission overhead, particularly in wireless environments with limited bandwidth. For swarms exceeding 30 UAVs, an extended 50-UAV evaluation and a side-by-side Zenoh benchmark are provided in Appendix C. The appendix details discovery traffic, success rate, and architectural implications at that scale, offering practical guidance for future deployments involving over 100 UAVs.

Figure 12a presents results from a scenario similar to Figure 11a, but with each UAV publishing five topics concurrently. As the number of publishers increased, Fast DDS dropped below an 80% success rate when more than four UAVs were active. Connext DDS and Cyclone DDS showed similar degradation beyond six and seven UAVs, respectively. In contrast, the proposed discovery mechanism maintained a success rate above 80% even in a 10-UAV setup, demonstrating reliable message delivery under multi-topic transmission conditions.

We also evaluated Zenoh [44], a recently introduced alternative to DDS, by analyzing its initial packet burst behavior in a 5-UAV scenario using the same topic configuration as in Figure 10a. Although Zenoh differs significantly from DDS in its protocol architecture, both systems broadcast discovery messages at startup. As shown in Figure 12b, Zenoh generated a sharp surge in per-second packet rate immediately following initialization, which imposed a sudden load on the network. In contrast, the proposed method staggered the discovery process over time, thereby reducing the initial traffic spike. These results reinforce that the proposed approach effectively minimizes discovery overhead, even when compared with modern alternatives like Zenoh. To clarify, Zenoh is not based on the DDS standard, and its discovery semantics and message flow fundamentally diverge from those of DDS middleware. Accordingly, the traffic distribution shown in Figure 8 does not reflect an identical discovery mechanism but is instead presented solely as a quantitative reference for total discovery traffic. Owing to these underlying protocol differences, Zenoh is treated as a supplementary benchmark rather than a direct DDS counterpart. A more comprehensive comparison with Zenoh is presented in Appendix C.

To further examine performance characteristics, we analyzed transmission behavior at both the medium access control (MAC) and Physical (PHY) layers using the NS-3 simulator. The MAC layer handles channel access, retransmission control, and collision avoidance. Within this layer, data frames are enqueued for transmission; however, if repeated backoffs or collisions cause the queue to overflow, the affected frames are discarded and counted as MAC drops. The PHY layer manages the actual transmission and reception of wireless signals. At this layer, packet loss may result from interference, signal attenuation, or physical collisions. If transmission fails and the retry limit is exceeded, the frame is classified as a PHY drop. To evaluate DDS performance, we defined communication metrics for both layers, as summarized in

Table 2. Symbols used DDS traffic counters.

Symbol	Definition
$N_{\mathrm{tx}}^{\mathrm{MAC}}$	Total MAC-layer transmission attempts (initial + retransmission)
$N_{\mathrm{drop}}^{\mathrm{MAC}}$	Frames discarded by the MAC
$N_{\mathrm{tx}}^{\mathrm{PHY}}$	Total PHY transmission attempts (initial + retransmission)
$N_{\mathrm{retx}}^{\mathrm{PHY}}$	Number of retransmitted PHY frames (a subset of $N_{\mathrm{tx}}^{\mathrm{PHY}}$)
$N_{\mathrm{rx}}^{\mathrm{PHY}}$	Total frames successfully received at the PHY layer
$N_{\mathrm{drop}}^{\mathrm{PHY}}$	Frames discarded by the PHY

.

To better understand the performance improvements achieved by the proposed discovery mechanism, we analyzed network-layer transmission behavior using the same configuration described in Figure 11a. The number of UAVs was incrementally varied at 5, 10, 15, and 20. We assessed effectiveness and identified potential bottlenecks in each DDS implementation by measuring three key metrics: the PHY drop density ($R_{\mathrm{drop}}^{\mathrm{PHY}}$), MAC drop ratio ($R_{\mathrm{drop}}^{\mathrm{MAC}}$), and PHY retransmission ratio ($R_{\mathrm{retx}}^{\mathrm{PHY}}$). These metrics were derived from the detailed traffic counters presented in

Table 3. Comparison of DDS traffic counters by node count.

DDS	Nodes	MAC Layer Counters			PHY Layer Counters
		$\overline{N_{\mathrm{tx}}^{\mathrm{MAC}}}$	$\overline{N_{\mathrm{drop}}^{\mathrm{MAC}}}$		$\overline{N_{\mathrm{tx}}^{\mathrm{PHY}}}$	$\overline{N_{\mathrm{retx}}^{\mathrm{PHY}}}$	$\overline{N_{\mathrm{rx}}^{\mathrm{PHY}}}$	$\overline{N_{\mathrm{drop}}^{\mathrm{PHY}}}$
ConnextDDS	5	1765.0	382.2		5228.5	3845.7	20,729.3	184.7
	10	66,195.7	38,136.0		105,955.0	77,895.3	944,429.3	9165.7
	15	175,573.0	132,392.7		144,926.7	101,746.3	1,989,258.3	39,715.0
	20	269,827.0	231,970.7		115,272.3	77,416.0	2,121,877.0	68,297.3
CycloneDDS	5	1027.0	20.5		3601.5	2595.0	13,937.5	468.5
	10	4787.0	817.0		14,674.0	10,704.0	127,086.7	4979.3
	15	24,188.7	9612.0		43,435.7	28,859.0	582,037.0	26,062.3
	20	90,054.3	49,879.7		91,508.3	51,333.7	1,659,114.7	79,543.7
FastDDS	5	2818.0	871.5		8392.7	6446.2	33,416.8	153.8
	10	196,838.3	170,144.7		114,511.3	87,817.7	1,025,030.3	5571.7
	15	396,507.0	362,680.3		135,503.7	101,677.0	1,873,223.3	23,828.0
	20	485,902.0	456,977.7		104,396.7	75,472.3	1,932,302.0	51,234.7
Ours	5	573.3	12.3		1979.3	1418.3	7816.0	101.3
	10	2113.3	275.7		7876.3	6038.7	70,094.0	793.0
	15	5232.0	1351.7		16,802.3	12,922.0	232,490.3	2742.3
	20	8765.7	3221.0		23,688.3	18,143.7	443,523.3	6555.0

, which summarizes MAC and PHY layer statistics for each DDS implementation as the number of UAVs increases. Together, these metrics offer a layered evaluation of wireless performance—capturing LQ at the physical (PHY) level and congestion resilience at the MAC level.

Notably, in some DDS implementations, an inverse trend was observed: $N_{\mathrm{tx}}^{\mathrm{PHY}}$ and $N_{\mathrm{retx}}^{\mathrm{PHY}}$ decreased despite an increase in the number of nodes. This behavior was attributed to severe MAC-layer congestion, which caused a substantial number of packets to be dropped before reaching the physical layer.

The PHY drop density ($R_{\mathrm{drop}}^{\mathrm{PHY}}$), defined in Equation (4), represents the proportion of total PHY-layer transmission attempts that failed to result in successful reception. This metric isolates physical-layer influences, such as signal attenuation, external interference, and receiver sensitivity, independent of MAC-layer mechanisms like CSMA/CA. As each transmitted frame may be independently dropped by up to $(Nodes-1)$ receiving nodes, $R_{\mathrm{drop}}^{\mathrm{PHY}}$ is not theoretically bounded by $100\%$.

$$R_{\mathrm{drop}}^{\mathrm{PHY}}={\displaystyle \frac{N_{\mathrm{drop}}^{\mathrm{PHY}}}{N_{\mathrm{tx}}^{\mathrm{PHY}}}}\times 100$$

(4)

The MAC drop ratio ($R_{\mathrm{drop}}^{\mathrm{MAC}}$), defined in Equation (5), represents the percentage of frames enqueued at the MAC layer that failed to be transmitted due to queue overflow or exceeding the retry limit. As node density increases, contention among devices intensifies, leading to repeated backoffs and eventual frame drops. This ratio functions as a direct indicator of congestion-induced frame loss.

$$R_{\mathrm{drop}}^{\mathrm{MAC}}={\displaystyle \frac{N_{\mathrm{drop}}^{\mathrm{MAC}}}{N_{\mathrm{tx}}^{\mathrm{MAC}}}}\times 100$$

(5)

The PHY retransmission ratio ($R_{\mathrm{retx}}^{\mathrm{PHY}}$), defined in Equation (6), indicates the percentage of total PHY transmission attempts that involved retransmissions. These retransmissions are initiated at the MAC layer when a collision is anticipated, and the PHY layer carries out the repeated transmission. A high retransmission ratio reflects increased channel occupancy, which can lead to higher latency and reduced output.

$$R_{\mathrm{retx}}^{\mathrm{PHY}}={\displaystyle \frac{N_{\mathrm{retx}}^{\mathrm{PHY}}}{N_{\mathrm{tx}}^{\mathrm{PHY}}}}\times 100$$

(6)

As shown in

Table 4. Comparison of DDS performance metrics across varying node count.

DDS	Nodes	${\mathit{R}}_{\mathbf{drop}}^{\mathbf{PHY}}$ (%)	${\mathit{R}}_{\mathbf{drop}}^{\mathbf{MAC}}$ (%)	${\mathit{R}}_{\mathbf{retx}}^{\mathbf{PHY}}$ (%)
ConnextDDS	5	3.52	21.22	73.53
	10	8.60	57.47	73.50
	15	27.43	75.40	70.20
	20	59.27	85.93	67.17
CycloneDDS	5	13.22	1.98	72.03
	10	33.90	17.07	72.93
	15	59.90	39.70	66.37
	20	86.90	55.40	56.07
FastDDS	5	1.82	30.42	76.82
	10	4.87	86.47	76.70
	15	17.60	91.43	75.03
	20	49.10	94.00	72.30
Ours	5	5.17	2.20	71.63
	10	10.07	13.03	76.67
	15	16.33	25.77	76.90
	20	27.67	36.70	76.57

Bold values indicate the best-performing DDS implementation at each node count.

, each DDS implementation exhibits distinct patterns of packet loss and retransmission across protocol layers as the number of UAVs increases from 5 to 20. Although both PHY and MAC drop rates generally increased with swarm size, the primary performance bottleneck differed among DDS vendors.

For Connext DDS and Fast DDS, MAC drop rates rose sharply to 75% and 94%, respectively, as the number of UAVs increased from 15 to 20. This degradation was primarily due to queue saturation and retry-limit violations at the MAC layer. Although PHY retransmission ratios plateaued between 67 and 77%, additional retries failed to improve delivery success.

Cyclone DDS recorded a PHY drop rate of up to 87% at 20 UAVs. This indicates that physical layer impairments—such as interference and attenuation—were the dominant cause of communication failure, while MAC-level losses remained relatively low. Despite this, overall success rates declined sharply due to excessive PHY-layer packet loss.

In contrast, the proposed discovery mechanism demonstrated the lowest packet loss rates under identical test conditions, with a PHY drop of 28% and a MAC drop of 37%. Although its retransmission ratio (76–77%) was similar to that of other vendors, the proposed method substantially reduced effective loss and maintained stable transmission, even in large-scale swarm scenarios.

Unlike other DDS implementations that encounter bottlenecks at both the MAC and PHY layers, the proposed approach mitigates losses across both layers, sustaining high communication efficiency as the number of UAVs increases.

5.2. PX4-ROS2 Simulation Experiments

To validate the proposed discovery algorithm in a PX4–ROS2-based UAV swarm environment, we conducted simulation experiments using Gazebo, as shown in Figure 13. Gazebo is a widely used and validated platform for simulating UAV operational scenarios. The PX4–ROS2 system communicated with Gazebo through the MAVLink protocol [45], and all UAVs operated using a host network interface.

To evaluate the effectiveness of the discovery mechanism, Gazebo was integrated with the NS-3 network simulator. UAVs were positioned linearly along the x-axis at 1 m intervals from a designated Commander UAV. RSSI values were configured to degrade with distance, allowing controlled variation in simulated packet loss.

Each UAV joined the DDS domain and transmitted a readiness signal to the Commander UAV. If the Commander UAV did not receive the signal within 30 s, the event was classified as a discovery failure.

As shown in Figure 14, the number of UAVs in the simulation was increased incrementally from 5 to 30 in step five. We measured the time each UAV required to receive the Takeoff command. The average reception time for each DDS implementation was plotted as a line graph, and results from 20 repeated trials were presented using boxplots to reflect statistical variance.

Although the proposed method introduces slight latency due to its distributed, network-aware architecture, it demonstrated consistent performance as the swarm size increased. In the 10-UAV configuration (Figure 14a), ConnextDDS and FastDDS exhibited greater communication delays than the proposed method, particularly for UAVs positioned farther from the commander, while CycloneDDS maintained relatively low latency. At 15 UAVs (Figure 14b), FastDDS failed to establish communication with several UAVs, and ConnextDDS showed increasing latency with distance. CycloneDDS began to exhibit slight delays but continued to outperform both FastDDS and ConnextDDS.

From the 20-UAV configuration (Figure 14c), CycloneDDS began to exhibit longer delays than the proposed method, and ConnextDDS experienced partial communication failures. In the 30-UAV scenario (Figure 14d), the performance gap widened further. ConnextDDS failed to communicate with all but one UAV (Node 2), and FastDDS exhibited similar failures. CycloneDDS maintained limited communication but showed delays ranging from 15 to 30 s. In contrast, the proposed scheme sustained communication with most UAVs and kept latencies below 15 s, demonstrating superior scalability and robustness in high-density deployments.

The communication success rate of the TakeOff command was evaluated as the number of participating UAVs increased. As shown in Figure 15, all DDS implementations except FastDDS achieved a success rate above 99% in the 10-UAV configuration. However, as the swarm size increased to 30 UAVs, the performance differences among implementations became more pronounced. In this high-density scenario, the proposed method sustained a success rate of 92.41%, while CycloneDDS, FastDDS, and ConnextDDS exhibited substantial declines. These findings indicate that the proposed distributed mechanism—leveraging network awareness and preloading discovery—provides greater communication stability and performance in wireless UAV swarm environments.

5.3. Real-World Experiments

To validate the proposed discovery algorithm in a real UAV environment, we constructed a PX4–ROS 2 system using Pixhawk flight controllers and Jetson Nano computers and conducted experiments with 10 UAVs. The UAV hardware specifications are presented in Figure 16. The wireless network was established using an Iptime AX8008M access point, and each UAV was equipped with an Intel 8265NGW Wi-Fi module connected to the Jetson Nano. The DDS configuration in the PX4–ROS 2 system was kept identical to that of the simulation setup to ensure consistency.

As shown in Figure 17, the packet transmission trends observed in the real-world environment closely mirrored those seen in the simulation(Figure 10a). FastDDS exhibited a sharp spike in transmission rate during the discovery phase, while CycloneDDS showed a more moderate increase, although it still caused a noticeable traffic surge. In contrast, the proposed discovery method, which adapts to real-time network conditions, effectively avoided such bursts. However, due to its decentralized structure, the proposed method introduced slight delays in completing the discovery phase. ConnextDDS was excluded from the real-world experiments due to its commercial license restrictions.

Table 5. Peak traffic comparison across DDS vendors in real-world environments.

DDS Implementation	Peak Traffic (Packets/s)			Δ from Ours (%)
	Mean	Std. Dev.		Mean ↓	Std. Dev. ↓
FastDDS	2593.56	3948.21		97.64%	97.21%
CycloneDDS	852.89	1312.24		92.83%	91.60%
Ours	61.11	110.19		–	–

Bold values indicate the lowest (best) traffic metrics among all implementations. ↓ indicates the percentage reduction relative to our method.

summarizes the peak traffic metrics observed during the discovery phase. The proposed method significantly outperformed existing DDS implementations, reducing both peak mean and standard deviation by over 90% compared to FastDDS and CycloneDDS. These results highlight the effectiveness of the proposed discovery mechanism in mitigating excessive startup traffic.

6. Conclusions

This research proposes two complementary optimization strategies—network awareness and preloading—to mitigate congestion during the DDS discovery phase in swarm-UAV environments. Network awareness introduces a cross-layer mechanism that adjusts discovery timing based on real-time RSSI feedback from the physical layer, whereas preloading reduces redundant signaling by embedding EDP information into each UAV prior to deployment. Since both techniques are integrated within the existing DDS framework, the scheme controls discovery traffic without redesigning the core architecture or requiring additional hardware, offering a favorable trade-off between implementation cost and communication scalability.

Simulations and field tests demonstrate that the proposed mechanism reduces MAC-layer packet loss by more than 20% and suppresses initial traffic spikes by over 90%. Even with 30 UAVs communicating simultaneously, the data-reception success rate remained consistently above 80%. These results confirm the method’s effectiveness in enhancing the scalability and reliability of DDS-based wireless swarm operations.

The present evaluation covers three ROS 2-compatible DDS vendors—Fast DDS, Connext DDS, and Cyclone DDS—and swarms of up to 30 UAVs; larger deployments and alternative middleware remain to be validated. Future studies will, therefore, focus on ultra-large swarms, lightweight cryptographic handshakes (Appendix B) to thwart spoofing while preserving preloading efficiency, and heterogeneous wireless networks that combine multiple access technologies under dynamic link conditions such as network partitions or rapid RSSI fluctuations.