The goal of NETWIN is to leverage the synergism between artificial intelligence and networks in two directions:

  • design machine learning strategies that will guide the design of autonomous networks able to efficiently handle the complexity of the telecommunications of the future with minimum human intervention;
  • design network architectures able to promote the pervasive deployment of delay-sensitive and energy-constrained intelligent services.

The project promotes also a paradigm shift from conventional Shannon-based approach to a goal-oriented semantic framework that exploits suitable knowledge representation systems to make the exchange of information more efficient than conventional systems. The approach adapts the coding scheme to the task underlying the exchange of information and to the channel state, in order to find an optimal trade-off between rate and semantic distortion.

Special attention is posed on distributed machine learning strategies, running in the edge cloud, to provide delay-sensitive services with minimum energy consumption. The project includes the application of the developed novel methodologies to three use cases: intelligent robots exploiting extended perception capabilities enabled by the edge cloud, aerial-assisted intelligent services exploiting digital twins of the UAV, and virtual assistants based on AI-aided extended reality.

Netwin is parte of Spoke: 8 – Intelligent and Autonomous Systems

Project PI: Sergio Barbarossa

  • The first collaborative effort of NETWIN has been the identification of the main scenarios of interest and the relative architectures. The results of this activity have been collected in the first milestone, produced at the end of June 2023.
  • The second collaborative effort has been the development of three proofs-of-concept implementing some of the most innovative technologies developed during the project in real hardware: a) a moving robot exhibiting intelligent recognition and navigation capabilities thanks to fast computation offloading of computing in the edge cloud; b) an aerial service based on drones having sensing, processing and communication capabilities, capable of taking decisions on what is being observed; c) a virtual assistant system producing a hologram that interacts with a human exploiting a large language model properly trained to reduce hallucinations end tailored to the context. 
  • At the methodological level, NETWIN proposed novel architectures exploiting generative AI for semantic and goal-oriented communications, with the goal of identifying and transmitting only what is strictly relevant, at the receiver side, to trigger a generative model able to recover a semantically equivalent version of what has been transmitted, taking explicitly into account the goal of communication. 
  • NETWIN focused also on the use of AI to improve the efficiency of a telecommunication network, at the core and RAN level, focusing on the implementation of microservice applications in the edge cloud. 
  • NETWIN developed distributed learning mechanisms, to be implemented in the edge cloud, with the goal of enabling low latency services and improving network efficiency. 
  • NETWIN developed a digital twin emulating the behavior of a UAV-assisted intelligent service, able to complement the information provided by the UAV with the digital twin to improve the understanding of the context and the navigation capabilities. 
  • Innovative semantic communication architectures, based on deep generative models such as variational autoencoders and probabilistic diffusion models; the algorithms have been able to achieve image compression rates of 0.01 bit per pixel (BPP), with a perceptual quality comparable with state of the art image compression schemes, but at a much higher BPP
  • The implementation of microservice applications in the edge cloud using multi-cluster Kubernetes enabled very fast orchestration and migration of stateless services in the edge cloud, which made possible fast navigation of moving robots through computation offloading and service migration 
  • A new cloud network flow model for the optimization of microservice applications valid for arbitrary network topologies and able to accurately compute end-to-end service latency that provided better performance in terms of energy consumption-service delay and decision accuracy than state of the art methods 
  • Service placement and request routing including disaggregation of services and network functionalities 
  • Multi-agent reinforcement learning techniques for dynamic control policies to be used for delay-constrained routing of delay-sensitive applications 
  • Novel multi-sensor learning-based architectures for cooperative prediction and inference in partially unknown dynamical systems over sensor networks 
The first demonstrator has been implemented and tested. It consists of a mobile robot that is endowed with intelligent capabilities thanks to computation offloading from the robot to the edge cloud, exploiting seamless integration of multiple radio access technologies (millimeterwaves at 60 GHz WiFi and sub-6 GHz 5G) and cloud/edge 
resources. The testbed leverages Kubernetes, Istio service mesh, OpenFlow, public 5G networks, and IEEE 802.11ad mmWave (60 GHz) Wi-Fi access points. The architecture is validated through a use case in which a ground robot autonomously recognizes and follows a moving object by using an artificial intelligence-driven computer vision application. Computationally intensive navigation tasks are offloaded by the robot to microservice instances, which are executed on demand within cloud and edge data centers that the robot can exploit during its journey.

Papers:

Giuseppe Baruffa; Andrea Detti; Luca Rugini; Francesco Crocetti; Paolo Banelli; Gabriele Costante; Paolo Valigi, “AI-Driven Ground Robots: Mobile Edge Computing and mmWave Communications at Work”, IEEE Open Journal of the Communications Society, 2024 

Francesco Binucci; Paolo Banelli; Paolo Di Lorenzo; Sergio Barbarossa, “Opportunistic Information-Bottleneck for Goal-Oriented Feature Extraction and Communication”, IEEE Open Journal of the Communications Society, 2024 

Sergio Barbarossa; Danilo Comminiello; Eleonora Grassucci; Francesco Pezone; Stefania Sardellitti; Paolo Di Lorenzo, “Semantic Communications Based on Adaptive Generative Models and Information Bottleneck”, IEEE Communications Magazine, 2023 
Fastweb and Logogramma just joined the project to contribute to the development of the proof-of-concept on AI-aided extended reality services.
Program partners: Cascade calls partners:
  • Dr. Eleonora Grassucci submitted a proposal for a Special Session at ICASSP 2024, Seoul, South Korea, entitled “Generative Semantic Communication: How Generative Models Enhance Semantic Communications”, that has been accepted.
  • Prof. Antonia Tulino gave a plenary talk entitled “A service-driven network evolution: from communication, to content distribution, to ubiquitous computation’’, at the IEEE International Symposium on Information Theory, June 30th, 2023.
  • Prof. Sergio Barbarossa gave an Invited talk entitled "Semantic and Goal-Oriented Communications: From Adaptive Generative Models to Optimal Resource Allocation", at the Huawei Strategy and Technology Workshop, Munich, Oct. 26th, 2023.
The main challenge of Netwin project is to bring Artificial Intelligence (AI) algorithms within the network design, with a twofold objectives:
1) make the networks more efficient thanks to additional capabilities in predicting behaviors and then allocate networks resources, radio, computation and storage, in an intelligent way;
2) use pervasive deployment of radio access points to bring intelligent services as close as possible to the end-user in order to promote delay-critical applications.

The project is also promoting a paradigm shift towards next-generation (6G) networks, exploiting generative AI algorithms to design innovative semantic and goal-oriented communications, where the idea is to transmit only the data that are necessary at the receiver side to generate a message that is semantically equivalent to the transmitted one, while not being necessarily identical at the bit level. This new perspective provides many additional degrees of freedom that can be exploited to make the overall transmission system more efficient. The first proof-of-concept prototype that has been implemented within the first half of the project is illustrative of the capability of the new design bring intelligent services in a mobile robot, with negligible delays due to task offloading mechanisms, thanks to the seamless integration of multiple radio access technologies (multi-RAT) and cloud/edge resources, unifying distributed and heterogeneous computing and communication resources into a cohesive continuum system, tailored for mobile applications
Publications
  • Expected: >= 36 publications in 36 months
  • Accomplished: 77 (32 journal + 31 conference + 14 submissions)
  • Readiness: 100%
Joint Publications
  • Expected: >=12 (30% joint publications in 36 months)
  • Accomplished: 5
  • Readiness: 62.5% (joint collaborations have just started, joint publications will follow)
Talks/Communication events
  • Expected: 10 talks or event chairing/organizing within NETWIN activities in 36 months
  • Accomplished: > 10 (6 invited talks, 1 special session at IEEE ICASSP 2024, 2 talks at RESTART General Meetings, and 31 conference presentations)
  • Readiness: >100%
Demo/PoC
  • Expected: 3 PoCs expected by the end of the project
  • Accomplished: 3 prototypes running
  • Readiness: 100% (the activity is not over yet, but the three prototypes are already running and are expected to be finished by the end of the project)
Project Meetings
  • Expected: > 12 meetings
  • Accomplished: 6 plenary project meetings + several WP  biweekly meetings
  • Readiness:100%
Patents/Innovations
  • Expected: 4 items over 36 months
  • Accomplished: 0 items submitted to mission 7
  • Readiness: 0%
M1.1 Scenarios and Architectures
  • Intermediate report First report on identification of use cases
  • 30/06/23 achieved
M2.1 Semantic and goal-oriented communications
  • Intemediate report Definition of semantic network architectures, development of algorithms for joint semantic encoding/decoding using generative models
  • 31/12/23 achieved
M3.1 Intelligent services for communications
  • Intermediate report Development and test of preliminary algorithms aimed at introducing machine learning algorithms for network control
  • 31/01/24 achieved
M4.1 Distributed delay-sensitive intelligent services
  • Intermediate report Development and test of preliminary algorithms for distributed learning in the edge cloud
  • 28/02/24 achieved
  • Expected: 4
  • Accomplished: 4
  • Readiness level: 100%

Researchers involved: Before the cascade call, the project was composed of 6 partners, involving in total 37 researchers. After the cascade call, additional 8 partners have joined the project with 32 more researchers.

Collaboration proposals
Given its methodological imprint, NETWIN is fully open to collaborations and it has already identified the following collaborations with other two other RESTART projects: 6GWINET and SUPER.
In the first phase of the project, NETWIN has collaborated with Spoke 8 to develop a joint strategy for the acquisition of instrumentations to be used for the project, such as a private 5G network. Furthermore, NETWIN has worked with Spoke 8 to set up the cascade calls useful to include new partners having specific expertise on the technologies involved in NETWIN proofs of concept.

For any proposal of collaboration within the project please contact the project PI.

Netwin News: