Toward Experimentation-as-a-Service in 5G/6G: The Plaza6G Prototype for AI-Assisted Trials

T o ward Experimentation-as-a-Service in 5G/6G: The Plaza6G Prototype for AI-Assisted T rials Sergio Barrachina-Muñoz, Marc Carrascosa-Zamacois, Horacio Bleda, Umair Riaz, Y asir Maqsood, Xavier Calle, Selva Vía, Miquel Payaró, Josep Mangues-Bafalluy Centr e T ecnològic de T elecomunicacions de Catalunya (CTTC/CERCA) , Barcelona, Spain {sbarrachina, mcarrascosa, hbleda, uriaz, ymaqsood, xcalle, svia, mpayaro, jmangues}@cttc.cat Abstract —This paper pr esents Plaza6G, the first operational Experiment-as-a-Service (ExaS) platform unifying cloud r esources with next-generation wireless infrastructure. Developed at CTTC in Barcelona, Plaza6G integrates GPU-accelerated compute clus- ters, multiple 5G cores, both open-source (e.g., Fr ee5GC) and commercial (e.g., Cumucore), programmable RANs, and physical or emulated user equipment under unified orchestration. In Plaza6G, the experiment design r equires minimal expertise as it is expressed in natural language via a web portal or a REST API. The web portal and REST API are enhanced with a Large Language Model (LLM)-based assistant, which employs retrie val-augmented generation (RA G) f or up-to-date experiment knowledge and Low-Rank Adaptation (LoRA) f or continuous domain fine-tuning . Over -the-air (O T A) trials leverage a four - chamber anechoic facility and a dual-site outdoor 5G network operating in sub-6 GHz and mmW av e bands. Demonstrations include automated CI/CD integration with sub-ten-minute setup and interactive O T A testing under programmable propagation conditions. Machine-readable experiment descriptors ensure r e- producibility , while future work targets policy-aware orchestra- tion, safety validation, and federated testbed integration toward open, reproducible wireless experimentation. Index T erms —Experiment-as-a-service (ExaS), network au- tomation, wireless testbeds, 5G, 6G, LLM I . I N T RO D U C T I O N The increasing complexity of next-generation (xG) wire- less networks necessitates e xperimentation en vironments that transcend simulation and theoretical modeling. Cloud-nativ e architectures, software-defined networking, and programmable radio interfaces are fundamentally reshaping wireless service design and validation. Howe ver , integrating realistic radio con- ditions with scalable computing resources remains challeng- ing. Our previous work [1] built upon the Experimentation-as- a-Service (ExaS) paradigm [2], [3] enabling on-demand, re- producible, and automated experimentation by embedding xG testbeds directly into Continuous Integration and Continuous Deployment (CI/CD) pipelines. While that work established the vision and architectural principles, this paper presents the first operational realization of ExaS via the Plaza6G platform. Dev eloped at CTTC in Barcelona, Plaza6G implements ExaS by unifying cloud computing flexibility with real 5G and 6G experimental infrastructure (Fig. 1). The platform provisions bare-metal servers, virtual machines/instances, Ku- bernetes clusters, GPUs, and xG radio and core networks This work was partially funded by Spanish MINECO grants TSI-064100-2022-16/-2023-26 (Plaza6G/Plaza6G+) and Grant PID2021-126431OB-I00 (ANEMONE) funded by MCIN/AEI/ 10.13039/501100011033 and by ERDF A way of making Europe W eb portal ExaS Scheduler Plaza6G brain LLM Assistant RAG + LoRA User Layer Orchestration Layer Experiment Modalities Simulation ns3, Komondor Compute 3000 CPUs, 30 vGPUs Network 200 Gbps links Infrastructure xG radio UEs, O-RAN, mmW ave Experiment repository REST API user Emulation srsRAN, Simbox Storage 500 TB In-lab OT A anechoic chamber Outdoors private 5G network Fig. 1: Conceptual architecture of Plaza6G. through unified orchestration. Users interact via a web portal or API enhanced with a large-language-model (LLM)-based interface that interprets natural-language requests and automat- ically configures compute and wireless resources. This design enables users with minimal domain expertise to deploy com- plex experiments immediately , eliminating manual scripting and configuration barriers. Plaza6G represents, to our knowledge, the first platform merging cloud-scale resources with real wireless infrastructure under an AI-assisted ExaS model. Existing 5G/6G testbeds typically require manual intervention, present fragmented in- terfaces, and demand specialized expertise, constraining ac- cessibility and hindering reproducibility . Plaza6G addresses these limitations through integrated automation and intelligent orchestration. This paper demonstrates Plaza6G’ s current capabilities through two representati ve use cases. First, the API is used to launch in parallel three emulated 5G networks, enabling automated CI/CD network acceptance testing as part of a software developers pipeline, where applications under dev el- opment are validated concurrently across different 5G core im- plementations. Second, a real over-the-air (O T A) experiment is ex ecuted via the web portal, in which a user requests the deployment of a complete 5G network with a physical user equipment (UE) and a gNB. These concurrent experiments highlight Plaza6G’ s scalability , reproducibility , and multi- tenant isolation capabilities. I I . T H E P L A Z A 6 G P L A T F O R M A N D R E S O U R C E S A. Ar chitectur e and Experiment Modalities Building upon the Experimentation-as-a-Service (ExaS) model introduced in [1], Plaza6G represents its first opera- tional instantiationa multi-domain en vironment enabling auto- mated wireless experimentation across four distinct modalities with varying fidelity lev els. The platform adopts a three-layer architecture: the user layer provides web portal and REST API access augmented by an LLM-based interface interpreting intents into ex ecutable experiment graphs; the or chestration layer coordinates re- source allocation through polic y-dri ven scheduling; and the infrastructur e layer e xposes heterogeneous compute, network, and radio assets as composable services. User interaction combines graphical workflo w composition with natural-language processing via an LLM-based assistant. The LLM backend, deployed locally at CTTC, le verages retriev al-augmented generation (RA G) and Low-Rank Adap- tation (LoRA) [4] to improve technical dialogue accuracy and orchestration safety while maintaining low latency . Fig. 2 shows the portal where natural-language intents are trans- formed into executable workflo ws. Plaza6G supports four experimentation modalities under unified orchestration: (i) Simulation employs discrete-e vent simulators ( ns-3 [5], K omondor [6]) for lar ge-scale protocol ev aluation; (ii) Emulation instantiates virtualized protocol stacks ( UERANSIM or srsRAN UE/gNB) for rapid multi- configuration benchmarking; (iii) In-lab integrates physical equipment (e.g., Amarisoft Callbox , commercial UEs) within a four-chamber anechoic facility for repeatable O T A validation; and (iv) Outdoors le verages a dual-site outdoor 5G network (sub-6 GHz, mmW av e) for end-to-end trials under realistic conditions. These modalities enable progressi ve validation from simulation through field deplo yment within the same framew ork. B. Infrastructur e and Resources Plaza6G infrastructure spans three integrated technological domains supporting all experiment modalities. The compute domain provides GPU-accelerated clusters hosting virtual ma- chines and Kubernetes workloads with support for edge-cloud continuum deployment. Current capacity exceeds 3,000 CPU cores, 30 vGPUs (NVIDIA L40S), and approximately 500 TB of storage, enabling concurrent execution of simulation, em- ulation, and virtualized network function workloads. The net- work domain offers multiple 5G core implementations such as F r ee5GC , Open5GS , or Cumucore , supporting end-to-end net- work slicing and service isolation across all experiment types. The r adio domain comprises programmable RAN platforms including Amarisoft Callbox , O-RAN, and srsRAN, alongside both emulated user equipment ( UERANSIM , Amarisoft Sim- box ) and physical devices (commercial Android smartphones). Controlled O T A testing le verages the four-chamber anechoic facility ensuring repeatable radio propagation conditions for in-lab experiments, while outdoor infrastructure supports field trials as described belo w . Fig. 2: Plaza6G web portal with the LLM assistant. Fig. 3: Coverage of the Plaza6G outdoor private 5G network at the PMT campus, operating across sub-6 GHz and mmW a ve bands. The color scale indicates recei ved signal power , from yellow (strong) to dark blue (weak). Plaza6G extends laboratory capabilities through a dual-site outdoor 5G network deployed across rooftop installations at the P ar c Mediterrani de la T ecnologia campus. Operating in both sub-6 GHz and mmW a ve bands, this network sup- ports configuration with open-source or commercial 5G cores, enabling field e xperimentation under realistic propagation, interference, and mobility conditions. In particular , the two outdoor radio sites can also be re-configured so that they belong to two independent 5G networks to test, e.g., roaming scenarios. Fig. 3 presents measured cov erage, demonstrating stable connectivity across the campus area served by the rooftop antenna deployment. C. P ositioning Against Existing T estbeds Large-scale European initiati ves including 5GENESIS [7], 5G-VINNI [8], VIT AL-5G [9], and 5G-EVE [10] ha ve estab- lished comprehensi ve v alidation infrastructures for 5G tech- nologies. Recent efforts such as 6G-SANDBO X [11] tar get early 6G experimentation, while J oiner [12] federates 11 UK testbeds with automation capabilities, and the IEEE 5G/6G In- nov ation T estbed pursues end-to-end 3GPP-compliant CI/CD integration (currently under dev elopment). Ho wev er , these platforms predominantly rely on predefined workflo ws requir- ing manual configuration, constraining both accessibility and automation potential. Plaza6G differentiates itself through three key innov a- tions: (i) unified orchestration spanning simulation, emulation, controlled in-lab, and field e xperiment modalities within a single platform, (ii) AI-assisted zero-touch experimentation accessible to users without domain-specific scripting expertise, and (iii) native support for the Experiment-as-Code (ExaC) paradigm [2], [3], [13], enabling fully reproducible, version- controlled experiment definitions deployable via natural lan- guage or programmatic APIs. This combination of elements positions Plaza6G as a next-generation platform bridging the gap between cloud elasticity and realistic wireless experimen- tation across the complete validation spectrum. I I I . D E M O N S T R A T I O N U S E C A S E S T o validate Plaza6Gs operational capabilities, two represen- tativ e scenarios are presented: (i) automated API-driv en vali- dation integrated with external CI/CD pipelines, and (ii) con- trolled O T A experimentation with physical 5G equipment. These use cases highlight Plaza6Gs ability to embed network acceptance testing within software workflo ws while also sup- porting reproducible O T A and field e xperiments across sub- 6 GHz and mmW a ve environments. A. Emulated Use Case: CI/CD Network Acceptance via API This scenario demonstrates Plaza6G as an automated valida- tion stage within continuous integration workflo ws. A CI/CD pipeline triggers experimentation via a REST API call ex- pressed in natural language. For instance, to validate applica- tion performance across different 5G core implementations, a user can submit: { "user_request" : "Deploy across three 5G cores (Open5GS, Free5GC, OAI-CN) and verify exceeds threshold for test approval." } For the sake of simplicity , in this demonstration, my_app is instantiated as iperf3 and my_kpi is mean throughput with an acceptance test threshold of 50 Mbps. The LLM backend interprets the request, identifies the application under test, target cores, and success criteria, then generates an experiment plan. The system returns one of three responses ( appr oved , clarification r equired , or denied ) based on resource availability and polic y constraints. CI/CD pipelines may proceed automat- ically upon approv al or incorporate human-in-the-loop revie w , depending on organizational trust policies. Each experiment instantiates a complete emulated 5G sys- tem comprising UERANSIM -based UE and gNB connected to one of the three 5G core implementations. A dedicated Data Network Name (DNN) virtual machine hosts the ap- plication server ( iperf3 -s ), while the emulated UE e xe- cutes iperf3 -c for 2 minutes, generating TCP and UDP traffic for benchmarking. The ExaS scheduler manages the complete lifecycle (resource allocation, UE/gNB/5GC/DNN instantiation, measurement collection, and teardown) through an asynchronous scheduler . All three experiments execute concurrently on isolated compute pools, with comprehensive telemetry such as throughput, latency , and CPU utilization automatically archived. 5G core and traffic protocol F5GC-TCP O5GS-TCP OAI-TCP F5GC-UDP O5GS-UDP OAI-UDP Throughput (Mbps) 0 50 100 150 200 250 300 350 400 DL UL Fig. 4: Throughput performance comparison for both UDP and TCP traffic across three 5G core implementations e xecuted concurrently via Plaza6G API. T ests duration of 120 seconds, with throughput logged e very second. Figure 4 depicts representati ve throughput distributions across the three 5G cores. All measured mean values exceed the specified 50 Mbps threshold, demonstrating consistent per- formance for user’ s app under test. From a De vOps perspec- tiv e, this enables “network acceptance testing" where CI/CD pipelines adv ance to staging or production only after minimum KPI thresholds are satisfied. Measured setup time remains below ten minutes per experiment, reducing configuration effort by over an order of magnitude compared to manual procedures. B. In-lab use case: Over-the-Air Contr olled Experiment The second scenario illustrates Plaza6Gs capability to au- tomate the provisioning of complex physical experimenta- tion en vironments while deliberately supporting human-in- the-loop experimentation. Unlike the CI/CD use case we discussed before, which tar gets fully automated and script- driv en validation, this scenario is designed for exploratory and interactiv e experimentation, where users manually access net- work elements and conduct measurements without predefined ex ecution scripts. The experiment provisioning phase is fully automated. Using the Plaza6G web portal and its LLM assistant, the user selects an e xperiment template that deploys a F ree5GC core, an Amarisoft Callbox gNB, and a commercial Android smartphone acting as UE. T emplates expose a curated set of commonly used 5G parameters (e.g., 100 MHz bandwidth, MIMO) to simplify initial configuration, while allowing users to ov erride or refine parameters either manually or through the LLM assistant. The gNB and UE are placed in separate chambers of the four-chamber anechoic facility , enabling programmable control of path loss, attenuation profiles, and interference conditions. Figure 5 shows the physical O T A experiment setup. The scheduler automatically provisions compute, core, and radio resources and establishes end-to-end connectivity , after which the en vironment is handed o ver to the user for interacti ve experimentation. Once the setup is complete, the user remotely accesses the UE via Vysor to manually ex ecute application-level tests, Fig. 5: Anechoic chamber with an Android UE and an Amarisoft Callbox for controlled OT A experimentation. install software, or explore network behavior under controlled radio conditions. By progressiv ely adjusting inter-chamber attenuation, users can interactiv ely study the impact of channel degradation on throughput, latenc y , and percei ved quality of experience. This mode of operation supports exploratory stud- ies such as adaptiv e streaming behavior , application robustness to radio impairments, and edgecloud service performance under varying link quality , without constraining the experiment to predefined workflows. Throughout the session, the LLM assistant can provide contextual information on current signal conditions, acti ve configurations, and runtime statistics in natural language, assisting users in interpreting observ ations while retaining full control over experiment ex ecution. I V . D I S C U S S I O N A N D O U T L O O K A. Repr oducibility and W orkflow T raceability All Plaza6G experiments follow a unified orchestration workflo w in which ev ery action, from initial resource pro- visioning to final teardown, is logged as a machine-readable descriptor capturing software versions, network topology , hardware identifiers, and configuration parameters. These de- scriptors are archi ved in a searchable experiment repository that supports version control and longitudinal performance tracking across multiple runs. In addition to local traceability , Plaza6G maintains a con- sistent naming and indexing scheme that links e xperiment descriptors, orchestration logs, and collected metrics. This structured organization enhances repeatability while allowing users to audit the full lifec ycle of an experiment, from creation to completion, through a unified web interface. T o foster transparency and community reproducibility , the detailed pro- cedure for replicating the emulated use case presented in Section III-A has been published on the pr otocols.io platform. This companion protocol demonstrates the same workflo w ex ecuted via the Plaza6G web portal rather than the API, providing additional insight into the graphical interface and user interaction process. 1 B. Conclusions and futur e work The demonstrations in Section III validate Plaza6G as a practical realization of the ExaS paradigm, unifying radio, core, and compute infrastructure under automated orchestra- tion. LLM-assisted interfaces and programmable APIs reduce the expertise required for wireless experimentation, enabling reproducible, concurrent trials with setup times belo w ten minutes. By coupling data-center automation with xG infras- tructure, Plaza6G turns network experimentation into an on- demand cloud service for developers, researchers, and vendors. Future work will extend Plaza6G along sev eral directions. (1) Policy-a ware orchestration will optimize scheduling for cost, energy , and radio resource usage. (2) Safety and vali- dation mechanisms will verify LLM-generated actions to en- sure correctness and reproducibility . (3) Planned LoRA-based fine-tuning will incrementally refine the LLM using selected orchestration logs and user dialogues, improving technical ac- curacy without full-model retraining. (4) Federated operation with external testbeds is en visioned to enable geographically distributed, multi-domain experiments; integration with ETSI OpenSlice [14] is under study to align Plaza6G with emerging open orchestration standards. R E F E R E N C E S [1] S. Barrachina-Muñoz, H. Bleda, M. Requena, S. Vía, M. Payaró, and J. Mangues-Baf alluy , “Experiment-as-a-Service in the Pipeline: Empowering CI/CD with xG Acceptance T esting, ” in W ir eless On- Demand Netw . Syst. and Serv . Conf. (WONS) . IEEE, 2025, pp. 1–4. [2] T . W . Edgar and T . R. Rice, “Experiment as a service, ” in 2017 IEEE Int. Symp. on T ech. for Homeland Security (HST) . IEEE, 2017. [3] M. Boniface et al. , “BonFIRE: A Multi-Cloud Experimentation-as-a- Service Ecosystem, ” in Building the Future Internet thr ough FIRE . Riv er Publishers, 2022, pp. 243–266. [4] E. J. Hu, Y . Shen, P . W allis, Z. Allen-Zhu, Y . Li, S. W ang, L. W ang, W . Chen et al. , “LoRA: Low-rank adaptation of large language models. ” ICLR , vol. 1, no. 2, p. 3, 2022. [5] A. Larrañaga et al. , “ An open-source implementation and v alidation of 5G NR configured grant for URLLC in ns-3 5G LEN A: A scheduling case study in industry 4.0 scenarios, ” Journal of Network and Computer Applications , vol. 215, p. 103638, 2023. [6] S. Barrachina-Muñoz, F . W ilhelmi, I. Selinis, and B. Bellalta, “Komon- dor: A wireless network simulator for next-generation high-density WLANs, ” in 2019 Wir eless Days (WD) . IEEE, 2019, pp. 1–8. [7] G. Xylouris et al. , “Experimentation and 5G KPI measurements in the 5GENESIS platforms, ” in Proceedings of the 1st W orkshop on 5G Measur ements, Modeling, and Use Cases , 2021, pp. 1–7. [8] A. J. Gonzalez, M. Xie, P . H. Lehne, and P . Grønsund, “Achieving high throughput and lo w latency with 5G: A real implementation experience, ” IEEE Communications Magazine , vol. 59, no. 10, pp. 84–90, 2021. [9] V . Charpentier et al. , “Utilizing the VIT AL-5G platform to advance 5G standalone integration with vertical industries, ” EURASIP Journal on W ireless Comm. and Networking , vol. 2025, no. 1, p. 27, 2025. [10] M. Gupta et al. , “The 5G EVE end-to-end 5G facility for extensiv e trials, ” in 2019 ICC workshops . IEEE, 2019, pp. 1–5. [11] P . Merino-Gomez, B. Garcia, C. Andreo, D. Artuñedo, and J. Macias, “On-demand T rial Networks ov er 6G-SANDBOX infrastructure, ” in 2024 Joint European Conference on Networks and Communications & 6G Summit (EuCNC/6G Summit) . IEEE, 2024, pp. 1021–1026. [12] S. Saunders, O. Dilley , and D. Simeonidou, “JOINER-NSF: UKs national facility for spectrum access innovation, ” in IEEE Intl. Symp. on Dynamic Spectrum Access Netw . (DySP AN) . IEEE, 2025, pp. 77–78. 1 The complete experimental protocol, including setup, ex ecution, and data collection steps, is available at https://www .protocols.io/view/ plaza6g- experiment- reproduction- protocol- use- case- a- dm6gpm6pjgzp/v1. [13] L. Aguilar et al. , “Experiments as Code and its application to VR studies in human-building interaction, ” Scientific Reports (Natur e P ortfolio) , vol. 14, no. 1, p. 9883, 2024. [14] ETSI, “ETSI OpenSlice: Open Source Platform for Service Orchestra- tion and Management, ” https://osl.etsi.org/, accessed: Oct. 24, 2025.