PaperlayerView exact revision on arXiv

Source-linked AI summary

Designing, Deployment and Field Testing of C2Stack for Networked Intelligent Software-Defined UAVs

Maxwell McManus, Zhaoxi Zhang, Sidharth Santhi Nivas, Yuqing Cui, Prem Sagar Pattanshetty Vasanth Kumar, Chenzhi Zhao, Nicholas Mastronarde, George Sklivanitis, Dimitris Pados, Elizabeth Serena Bentley, Zhangyu Guan

arXiv:2608.28854v1eess.SY

TL;DR

Real-world deployment and testing of networked UAV systems remains difficult, motivating an end-to-end framework for real-time control and evaluation. The paper presents C2Stack, a configurable protocol stack with modular control and programmable data planes, and demonstrates online optimization in dynamic aerial environments while documenting unstable convergence under high channel variance.

  • Problem

    Real-world deployment and testing of networked UAV systems is challenging, creating a need for practical frameworks that support experimental evaluation of data-driven control techniques.

  • Method

    C2Stack combines modular protocol layers, CNOS-based monitoring and control, and a programmable packet-processing suite for cross-layer algorithm development and networked UAV experimentation.

  • Results

    C2Stack enabled benchmark experiments in protocol self-configuration and interference source localization, while online trials showed that high A2A channel variance can produce inconsistent rewards and unstable convergence.

  • Takeaways & Limitations

    The experiments demonstrate the feasibility and value of real-world aerial network experimentation and provide a foundation for evaluating data-driven control algorithms in connected autonomous aerial networks.

Abstract

from arXiv · show

Unmanned Aerial Vehicles (UAVs) are emerging as critical enablers of next-generation wireless networking and autonomous systems. Despite their potential, deploying and testing networked UAV systems in real-world environments remains challenging, largely due to the absence of well-developed, end-to-end, ready-to-use protocol stacks. To fill this gap, we present C2Stack, a configurable protocol stack and experimental framework designed for real-time control, evaluation, and optimization of UAV networks. C2Stack incorporates a modular control plane, referred to as the~C2Stack Network Operating System (CNOS), alongside a programmable data plane that exposes APIs for cross-layer algorithm development, digital twin integration, and autonomous swarm control. In this article, we share our experience with the deployment and testing of C2Stack. We implemented C2Stack on a custom UAV swarm platform that integrates multiprocessor system-on-chip (MPSoC) radios with Intel NUC computing modules, enabling interoperability with various RF front ends. Field trials were conducted in both netted environments and large-scale outdoor test ranges, focusing on two representative use cases: (i) network utility maximization through online reinforcement learning, and (ii) collaborative interference source localization. The experiments demonstrate the feasibility of real-time, data-driven optimization in dynamic aerial environments, while also revealing practical challenges in field deployments of networked UAV systems, including power constraints, sensing limitations, and deployment logistics. We have made C2Stack source code available to the community under the MIT License, with the goal of establishing it as a foundational framework for experimental research on intelligent networked aerial systems.

I. INTRODUCTION

C2Stack addresses the lack of end-to-end infrastructure for rigorous UAV-network experimentation by combining a configurable protocol stack with a custom aerial platform and field deployments. The introduction frames real-world instability, limited existing testbeds, and systematic evaluation as central challenges.

  • Motivation: Real-world UAV-network deployment remains difficult because vibration, wind, and hardware factors destabilize flight and wireless-device performance.These effects vary with rotor configuration, controllers, calibration, processing capacity, and airframe design.
  • Research gap: Existing AI/ML and control efforts often rely on idealized network conditions, simplified communication models, or highly customized datasets.The paper argues that over-the-air field trials are needed to assess effectiveness in real networks.
  • System capabilities: The framework supports real-time UAV interaction and inter-UAV signaling for autonomous swarm control and wireless-network self-configuration.These capabilities connect deployed-node operation with data-driven control algorithms.
  • C2Stack contribution: C2Stack provides a configurable, modular protocol stack with APIs for integrating and evaluating user-defined AI/ML algorithms in real UAV-network scenarios.Its design emphasizes dynamic end-to-end network control and cross-domain interactions.
  • Aerial platform: The custom aerial platform supports C2Stack payloads and real-time monitoring and control between FPGA-based baseband processing and upper protocol layers.This platform is intended to expose capabilities and limitations of networked UAV systems experimentally.
  • Field testing: Field deployments evaluate dynamic protocol self-configuration and interference source estimation while documenting implementation complexity and swarm limitations.The experiments provide benchmark results and identify research challenges for future UAV-network experimentation.
  • Community release: C2Stack is positioned for community use through an MIT-licensed release containing source code, documentation, platform specifications, and experiment-replication materials.The planned repository is intended to support installation, integration, operation, and replication.

II. C2STACK FRAMEWORK DESIGN

C2Stack combines the CNOS control plane with a programmable protocol stack to support real-time monitoring, configuration, adaptive control, and user-defined experimentation in UAV networks.

  • Framework objectives: C2Stack enables real-time configuration and evaluation of network protocols, data-driven control algorithms, and adaptive modeling techniques.Its fine-grained control spans physical-layer waveforms through custom network applications.
  • Control plane: The CNOS control plane manages node- and network-level processes, including protocol adaptation, topology reconfiguration, situational awareness, and swarm optimization.Its process communication server coordinates internal, node-level, and network-level signaling through standardized interfaces.
  • Control plane: The CNOS GUI provides live network visualization, PPS state metrics, user control of protocol parameters, and monitoring of throughput, SINR, transmit power, and video.The interface renders GPS-based deployment information and supports operation through the backend.
  • Extensibility: The Autonomy Toolchain exposes APIs for user-defined optimization algorithms that synchronously adapt to PPS observations during operation.The C2-App Suite adds asynchronous integration of user-defined modules and feature sets, including interference detection and localization.

B. C2Stack Data Plane

The PPS is a modular, programmable data plane that coordinates protocol layers and enables dynamic cross-layer optimization while remaining adaptable to different communication devices and protocols.

  • PPS architecture: The PPS coordinates protocol layers for data forwarding and supports synchronous deployment of layer-specific processes during network operation.It is intended as a platform for CNOS-based monitoring and control rather than as a collection of newly developed individual protocols.
  • Layer architecture: The five-layer PPS architecture comprises APP, TSPT, NET, MAC, and PHY layers with distinct application, flow-control, routing, queueing, and waveform functions.The architecture is designed for compatibility with existing protocol stacks.
  • Layer architecture: C2Stack implements an interference-resistant wideband protocol for A2A and A2G links, while allowing relevant layers to be adapted or replaced with open-source ad hoc protocols.Independent object-oriented layer implementations provide modularity.
  • PHY layer: The PHY layer uses FPGA-based DSSS processing and SDPI-controlled adaptation of frequency, modulation, front-end gains, and code selection.The design supports real-time access to transmit and receive parameters through the Autonomy Toolchain, C2-Apps, or GUI.
  • PHY layer: The PHY frame contains a 128-symbol preamble, a 128-symbol training sequence, and a 2048-symbol dataframe divided into eight configurable 256-symbol subframes.Each subframe can include one to eight 16-symbol pilot sequences depending on modulation and channel conditions.

III. AERIAL PLATFORM DEVELOPMENT

The authors develop a modular medium-duty UAV platform carrying C2Stack payloads for large-scale aerial experimentation, with hardware integration designed for configurable radios and real-time control.

  • UAV platform: The aerial experimentation network uses custom medium-duty UAVs designed for stable flight, high payload capacity, and modular networking experiments.The platform’s structural design targets vibration reduction and payload support.
  • Deployment constraints: 4.08 kg is the total payload weight, including 3.17 kg of hardware and structural components plus a 0.91 kg custom PDU.With 20% payload redundancy, the module is suited to UAV platforms with capacity ≥5 kg; lighter platforms require hardware reductions and incur tradeoffs.
  • Payload implementation: The C2Stack payload combines an Intel NUC control computer with a Zynq UltraScale+ ZCU102 MPSoC for FPGA-based baseband processing.An AD-FMCOMMS3-EBZ front end supports transmission and reception from 70 MHz to 6 GHz with up to 56 MHz bandwidth.
  • Payload implementation: Only PHY-specific transmit and receive chains run on the MPSoC, while other protocol functions run on the NUC to support interoperability with diverse software-defined and FPGA-enabled radios.The design explicitly accommodates platforms such as Ettus USRP and RFSoC devices.
  • Payload implementation: The independent custom power delivery unit isolates the payload from the UAV power system to prevent resource contention during deployment.The complete module is mounted on the UAV undercarriage.
  • Experimental campaign: The experimental campaign evaluates AI/ML-driven UAV networking through reinforcement-learning QoS optimization and collaborative interference-source localization.The missions assess autonomous parameter adaptation and collaborative network-control capabilities in field experiments.

A. Mission 1: Link-level QoS Optimization

Mission 1 uses C2Stack for real-time RF parameter self-optimization, modeling each UAV's link-throughput maximization as a multi-armed bandit problem solved by lightweight reinforcement learning.

  • C2Stack dynamically reconfigures RF front-end parameters to enhance network QoS during operation.
  • Each UAV runs an independent RL agent that maximizes aggregate link throughput over the flight duration.
  • Agents select discrete RF front-end gain values and receive rewards based on observed aggregate throughput.
  • The MIX-MAB implementation separates learning into exploration and continuing phases rather than introducing a new machine-learning algorithm.
  • During exploration, agents update action probabilities and weights, then remove actions whose probabilities fall below the threshold η, typically set to 0.5.
  • Changes in network configuration or environmental dynamics can invalidate learned action distributions, so resetting probabilities enables reuse of prior weights in the new context.

B. Mission 2: Interference Localization

Mission 2 uses runtime network observations and collaborative UAV decisions to localize interference sources without dedicated RF sensing hardware, supporting interference-aware flight planning.

  • C2Stack enables collaborative interference-aware path planning by using observed network states at the flight controller.
  • The IDLS converts runtime network-metric fluctuations into interference-related insights for flight control.
  • Interference localization aggregates predicted source distances with detecting-node positions to enable triangulation.
  • A random forest regression model estimates the distance between each deployed UAV node and a local RF interference source.
  • The model's hyperparameters are coarsely tuned by randomized search and then fine-tuned through exhaustive grid search.

V. FIELD MEASUREMENTS AND ANALYSIS

C2Stack was evaluated in netted and open-air facilities spanning small-scale and nine-node swarm deployments. The trials exposed contrasting operational trade-offs in access, scale, and deployment logistics.

  • The evaluation covered the two missions using both a netted enclosure and a large open-air test facility.
  • The netted deployment was limited to two UAV nodes because of safety constraints associated with the platform size.
  • Adjacent storage and maintenance facilities reduced deployment time and enabled rapid device- and network-level evaluation in the netted enclosure.
  • Travel required multiple days, and recalibration after transport significantly increased network deployment time at the open-air facility.

B. Experimental Methodology

The experiments evaluate C2Stack through two field missions: online protocol self-configuration and offline/online interference-source localization. Results show dynamic aerial links, variable reinforcement-learning convergence, and dataset-dependent inference accuracy.

  • Experimental setup: The testbed used 2.4 GHz, two 20 MHz channels, 64-QAM, and mission-specific aerial deployments for C2Stack evaluation.Mission 1 used configured A2A/A2G links; Mission 2 placed an aerial node 5 m high and 20 m from a transmitting ground node.
  • Mission 2: interference localization: The interference-localization experiments used six datasets from a netted enclosure and a larger test facility for offline inference evaluation.Datasets A, B, C, and D came from the enclosure, while J and K came from the larger facility.
  • Initial measurements: Throughput in both A2A and A2G links was highly dynamic, even under static deployments, requiring time-varying network models.The measurements challenge assumptions that aerial channels can be adequately represented by standard noise distributions.
  • Mission 1: protocol self-configuration: MIX-MAB identified favorable gain settings, but A2A channel variance produced inconsistent rewards and unstable online convergence across nodes and trials.Observed selections included g_k = 45 dB, 40 dB, 60 dB, and 65 dB, with some agents unable to remove sub-optimal actions within the measurement horizon.
  • Mission 1: protocol self-configuration: Expanding the action space from K = 7 to K = 12 produced either persistent uncertainty or delayed convergence, depending on channel conditions.One trial identified g_k = 40 dB as optimal, while another exhibited chaotic learning caused by inconsistent observed rewards.
  • Mission 2: interference localization: Offline IDLS inference matched online operation because historical datasets were streamed in the same format as real-time network data, with accuracy varying by training data and distance.Additional training data improved prediction at d_m(τ) = 40 m in one case but increased variance elsewhere; larger distances reduced accuracy in another.

VI. LESSONS LEARNED AND CHALLENGES

Field testing exposed practical constraints affecting C2Stack’s deployment area, sensing, endurance, weather tolerance, and implementation performance. These limitations motivate RF-sensing improvements, longer flight times, and further benchmarking.

  • Deployment area: The prototype requires large open spaces, while urban restrictions and FAA regulations constrain swarm flight-test locations.These requirements complicate scheduling and the evaluation of swarm navigation policies.
  • Sensing capability: The current PHY design lacks in-situ RF sensing, preventing reliable spectral profiling and adaptive protocol configuration from real-time environmental measurements.The limitation directly affects environmental awareness during deployment.
  • Flight duration: Battery power and payload weight restricted experiments to approximately 15-minute flights, forcing interruptions for landing and battery replacement.Tethering can provide power but reduces deployment flexibility and adds payload weight that alters flight performance.
  • Weather conditions: Weather windows reduce opportunities for extended testing because available waterproof UAVs lack the payload capacity and features required by the experimental platform.The limitation affects scheduling rather than only communications performance.
  • Software-defined protocol stack: Python-based implementation accelerates development and modularity but may introduce delays that limit high-throughput, low-latency, or strict-timing experiments.The authors identify further benchmarking and hardware redesign as future directions.

A. Research Challenges

The paper identifies research challenges spanning algorithm design, experimental scenario design, and UAV hardware. These challenges arise from limited flight endurance, complex field logistics, weather, sensing needs, and airframe effects on networking.

  • Algorithm Design: Data-driven algorithms may require flight times that exceed UAV battery capabilities, while environmental factors are often omitted from power-aware designs.The discussion points toward adaptive strategies using RF observations and resilient learning methods.
  • Scenario Design: Large-scale, reproducible swarm experiments require service contracts, legal agreements, transport logistics, and coordination across facilities and weather conditions.Existing platforms may not support flexible path-planning or network-control scenarios.
  • Hardware Design: UAV airframes introduce blockage, vibration, and instability that can degrade network performance and create operational-safety risks.Hardware design must balance stability, resilience, payload capacity, and networking functionality.

VII. CONCLUSIONS

C2Stack accelerates implementation and evaluation of data-driven control algorithms for connected autonomous aerial networks. Two benchmark experiments demonstrate the feasibility and value of real-world aerial-network experimentation while motivating more capable platforms.

  • Conclusion: C2Stack supports rapid implementation and evaluation of data-driven control algorithms in connected autonomous aerial networks.The framework is presented as a swarm control platform for experimental research.
  • Conclusion: Protocol self-configuration and interference-source localization served as benchmark experiments demonstrating real-world aerial-network experimentation.The conclusion links these experiments to the need for practical validation of data-driven networking approaches.
  • Conclusion: The study outlines research challenges for making UAV-network experimentation platforms more capable, accessible, and flexible.This conclusion follows from the deployment lessons and identified limitations.

APPENDIX

C2Stack is benchmarked for packet-level latency and network scalability, combining component-level latency measurements with analytical and simulated control-plane traffic analysis. The results quantify end-to-end delay, per-hop overhead, and the tradeoff between network size and control-signal reliability.

  • Performance Metrics: C2Stack measures packet-level latency and scalability as reference benchmarks for future framework experiments.Packet-level latency is measured from the transmitter application layer to the receiver application layer; scalability uses achievable control-signal PDR as network size increases.
  • System Latency: Latency characterization separates PPS, SDPI forwarding, and FPGA processing across transmit and receive paths.PPS covers protocol-stack processing, SDPI covers NUC-to-FPGA DMA transfer, and FPGA processing covers waveform generation and transmission.
  • System Latency: 8.12 ms average end-to-end frame latency is achieved for single-hop links, increasing by roughly 6.21 ms per additional hop.The reported sum includes PPS, SDPI, and FPGA processes along both transmit and receive paths.
  • Network Scalability: Scalability analysis combines an analytical control-plane signaling model with a lightweight discrete network traffic simulator.The simulator estimates channel saturation as network size N increases and is intended for open-source release with C2Stack.
  • Network Scalability: Larger and less frequent control-signal intervals generally yield higher PDR as network size increases, while overlapping transmissions produce the lowest PDR.The analysis models an ad hoc N-node FDD network, randomizes inter-signal waiting within user-configured intervals, and estimates PDR under collision losses.
  • Network Scalability: In the initial 9-node swarm, both discovery and state-signaling intervals achieve PDR > 0.95.The simulator is used to estimate suitable control-signal intervals for different network sizes and PDR tolerances.
Loading 2608.28854v1…