Source-linked AI summary
Opportunities of Federated Learning in Connected, Cooperative and Automated Industrial Systems
Stefano Savazzi, Monica Nicoli, Mehdi Bennis, Sanaz Kianoush, Luca Barbieri
TL;DR
Mission-critical autonomous industrial systems need fast, communication-efficient distributed learning under URLLC, but decentralized FL must address changing environments, heterogeneous devices, and network constraints. The article examines FL integrated with sensing-decision-action loops and discusses cooperative driving and collaborative robotics. In the robotics scenario, diffusion FL outperforms ego learning for sparse non-IID data, while incurring higher communication overhead than gossip.
Problem
Mission-critical networked industrial systems need distributed learning that remains effective across heterogeneous devices, changing environments, and intermittent or weakly connected networks.
Method
The article analyzes decentralized FL integrated into sensing-decision-action loops, with applications in cooperative automated driving and collaborative robotics.
Results
Diffusion FL gives smaller loss than ego learning after 39 s for robots with 3% non-IID data, but requires 220 kB per round versus gossip's 92 kB.
Takeaways & Limitations
Decentralized FL is promising for cooperative automated driving and collaborative robotics, where agents learn through model exchange rather than centralized raw-data fusion.
Abstract
from arXiv · showhide
Next-generation autonomous and networked industrial systems (i.e., robots, vehicles, drones) have driven advances in ultra-reliable, low latency communications (URLLC) and computing. These networked multi-agent systems require fast, communication-efficient and distributed machine learning (ML) to provide mission critical control functionalities. Distributed ML techniques, including federated learning (FL), represent a mushrooming multidisciplinary research area weaving in sensing, communication and learning. FL enables continual model training in distributed wireless systems: rather than fusing raw data samples at a centralized server, FL leverages a cooperative fusion approach where networked agents, connected via URLLC, act as distributed learners that periodically exchange their locally trained model parameters. This article explores emerging opportunities of FL for the next-generation networked industrial systems. Open problems are discussed, focusing on cooperative driving in connected automated vehicles and collaborative robotics in smart manufacturing.
I. INTRODUCTION
Networked industrial systems combine autonomous machines, large-scale ML, and URLLC requirements, motivating distributed learning that integrates sensing, communication, and decision operations. The article examines FL opportunities for cooperative vehicles and industrial robots, including decentralized model exchange without sending raw training data.
- Networked intelligent machines require distributed ML integrated with sensing, communication, and decision operations for collaborative tasks involving safety, reliability, scalability, and latency.
- FL trains shared models by exchanging parameters among interconnected devices instead of sending local training data to a server.Model parameters include DNN weights and biases.
- Federated averaging alternates local SGD-based model optimization with communication rounds for learning a global model.
- Fully decentralized FL addresses server-client robustness and scalability concerns by having devices exchange local model parameters with neighbors and fuse received updates until convergence.
- The article studies FL as part of autonomous industrial systems' sensing-decision-action loop, focusing on cooperative automated vehicles and collaborative industrial robotics.A practical usage scenario presents FL as promising when supported by URLLC communications.
II. SENSING, DECENTRALIZED LEARNING AND COMMUNICATION CO-DESIGN
Decentralized FL places model training within each machine's sensing-decision-action loop and uses local cooperation rather than centralized raw-data movement. Its central design challenges are reducing communication, balancing local and collaborative learning, and continually adapting over URLLC links.
- Decentralized FL integrates ML model training into the sensing-decision-action loop of connected machines performing collaborative tasks.
- Collaborative learning lets networked machines augment local models by sharing ego knowledge without centralized orchestration.
- Decentralized FL favors sparse model-parameter exchange over server-client architectures and movement of large raw datasets across D2D links.
- The learning process must balance opportunistic local training with collaborative use of neighbors' experience to steer convergence.
- FL must support continuous learning and retraining over URLLC links as data distributions, environments, processes, or situations change.
III. DECENTRALIZED FL: EMERGING TRENDS
Decentralized FL exchanges local models over D2D networks while local devices minimize their own losses. DSGD, gossip, consensus, and diffusion regulate cooperation and convergence toward a global model.
- Decentralized FL alternates local model exchange with on-device loss minimization to promote convergence toward a global model minimizing aggregated local losses.
- Fully distributed implementations use DSGD over arbitrary connectivity graphs, with convergence guaranteed under strong convexity and doubly stochastic adjacency assumptions.
- Gossip, consensus, and diffusion algorithms regulate mutual model exchange among decentralized learners.
A. Consensus, diffusion and gradient negotiations
Consensus and diffusion methods exchange or negotiate model updates among selected partners, with update design shaping convergence behavior. Distributed ledgers can validate updates, but robustness against poisoning and adversarial manipulation remains unresolved.
- Consensus, diffusion and gradient negotiations: Consensus approaches sequentially average exchanged local model parameters, using random, time-varying, or optimized partner selection.
- Consensus, diffusion and gradient negotiations: Update strategies trade off convergence speed and model accuracy, while gossipgrad typically connects each cooperating agent to two other nodes.
- Consensus, diffusion and gradient negotiations: Diffusion adds gradient negotiation about neighbor-model adjustments, virtually expands local training data, and can boost convergence compared with gossip.
- Consensus, diffusion and gradient negotiations: Distributed ledgers can validate model updates through steps such as Proof-of-Work, but robust defenses against data poisoning and adversarial manipulation remain an open problem.
B. Improving communication efficiency
Communication-efficient decentralized FL must align reliable, low-latency wireless exchange with model size, access strategy, and on-device computation. Digital, analog, and hybrid approaches offer different ways to reduce communication overhead and support distributed optimization.
- Communication requirements: Model transmissions require packet error rates down to 10^-8 and transmission time intervals targeting 5 ms or below.These requirements prevent retransmissions from slowing convergence and align communication with SGD and model-adaptation times.
- Diffusion strategies: Diffusion adds gradient negotiations before computing rounds, exchanging how neighbor models should be adjusted and potentially accelerating convergence compared with gossip.
- Digital communication: Quantization, sparsification, and distillation reduce the communication overhead of digital FL, which uses scheduled half-duplex wireless access.
- Analog communication: Analog FL uses wireless signal superposition to average simultaneously transmitted neighbor models without time-scheduled access.Sparse uncoded or uncompressed parameters can be recovered with Lasso-type methods.
- Resource allocation: Resource allocation and medium-access scheduling trade on-device SGD computation against wireless parameter exchange and can prioritize devices with higher-quality data.
C. Emerging research challenges
Emerging decentralized FL research must adapt communication and learning to changing connectivity, heterogeneous tasks, and application-specific industrial systems. Key directions include jointly learning network cooperation and models while comparing consensus with opportunistic behavior.
- Communication and graph dynamics: Current FL designs generally overlook network-graph dynamics, intermittent links, and weakly connected components.Future work should balance local data collection, model adaptation, and cooperation among selected agents.
- Adaptive communication: Digital, analog, or hybrid implementations should be selected according to mobility, model size, computation and bandwidth resources, and connectivity quality.
- Adaptive communication: Learning the network graph or neighbor-selection policy jointly with the ML model is identified as a promising direction.
- Heterogeneous industrial tasks: Heterogeneous devices performing related tasks motivate multitask learning and investigation of when consensus better steers convergence than opportunistic or ego behavior.The analysis must also account for distributed-computation cost, federated-data distribution, and URLLC network design.
- Application domains: The paper highlights FL opportunities for mission-critical control in cooperative automated driving and collaborative robotics.
IV. DISTRIBUTED INTELLIGENCE FOR COOPERATIVE AUTOMATED DRIVING
Cooperative automated driving uses distributed learning over V2X networks to combine vehicle sensing while meeting time-critical control requirements. Decentralized FL exchanges model parameters rather than raw data and can learn common models for cooperative perception.
- Cooperative perception: Cooperative automated vehicles share environmental maps through V2X URLLC to extend the range and resolution of radar, camera, and lidar sensors.
- System architecture: Time-critical driving functions require computation closer to the vehicles, with part of the cooperative processing performed locally rather than solely at the mobile edge cloud.
- Decentralized learning: Decentralized FL can exchange smaller models over 3 ms, highly reliable V2X connectivity, reducing learning time by transmitting model parameters instead of raw data.
- Cooperative perception: In cooperative perception, vehicles fuse GNSS, IMU, lidar, camera, and radar data to improve location sensing accuracy and enable real-time responses.
- Decentralized learning: Decentralized FL can learn a common model for associating and fusing locally collected raw data while limiting V2X exchange to model parameters.Unlabeled imaging-measurement association, large data volumes, and computational complexity remain challenges.
V. DECENTRALIZED FL FOR COBOTS
Collaborative robots create a decentralized FL setting where learning must track changing workflows and support human-robot coexistence. The paper illustrates this setting with continuous operator-position classification and gossip-versus-diffusion evaluation over a robot D2D network.
- Collaborative robotics: Cobots operate with humans in shared, fenceless workspaces, making collaborative manufacturing a challenging setting for decentralized FL.
- Evaluation setup: The evaluation compares gossip and diffusion FL loss over time in a 15-robot D2D network using IID and non-IID federated datasets.The setup uses 8% and 3% of the full training dataset, respectively, with TTI=3 ms and BLER=10^-9.
- Continual learning: Robotic FL systems must support continual learning over URLLC links as workflow changes alter data dynamics.
- Heterogeneous tasks: Robots may perform distinct but strongly related tasks, requiring decentralized learning tools to account for task relationships.
- Mission-critical control: The example classifies human operators into six regions using real-time distance and direction-of-arrival estimates to guide robotic-manipulator motion.
VI. RESULTS AND DISCUSSIONS
The experiments compare decentralized FL with opportunistic learning for robotic agents over D2D networks, emphasizing validation loss, convergence time, and communication overhead. Diffusion performs best with scarce, non-IID data, while gossip uses less communication and opportunistic learning converges quickly but can retain high loss.
- Experimental setup: Robots either formed 15-agent FL groups over D2D networks or learned opportunistically from local data alone.The comparison evaluates cooperation versus disabling the D2D interface.
- Evaluation: The simulations measured average validation loss over time to assess consensus-round duration for real-time implementation.Gossip and diffusion were evaluated on a regular D2D network with robots having two neighbors each.
- Results: Diffusion FL improved training for robots with only 3% data and non-IID examples, reaching lower loss than ego learning after 39 s.Its gradient exchanges require 220 kB per round, and convergence time increases with network size.
- Results: Gossip required 92 kB per round, but was less effective with non-IID data and improved ego approaches after 23 s with IID data.This lower communication requirement comes with reduced effectiveness under heterogeneous data.
- Trade-offs: Opportunistic learning converged quickly without D2D communication but experienced large loss, making it viable mainly with sufficient data and specialized tasks.The setting considered URLLC-aligned frames with 1 kB payloads, 3 ms TTI, and BLER of 10^-9.
VII. CONCLUSIONS AND FUTURE DIRECTIONS
The article identifies decentralized FL over D2D communications as an opportunity for networked industrial systems and discusses applications and open problems in manufacturing and automotive settings. Continuous retraining supports changing environments but creates networking co-design demands, while future work targets heterogeneous cooperating devices and URLLC-dependent behavior.
- Conclusions: The article explores FL opportunities and applications in networked automated industrial systems underpinned by D2D wireless communications.Its open problems focus on manufacturing and automotive B5G verticals.
- Conclusions: Decentralized FL integrates cooperative model learning into each automated entity’s sensing-decision-action loop to improve knowledge discovery.The paper frames this integration within networked industrial systems.
- Conclusions: Continuous learning and retraining for changing environments significantly affect networking co-design, including model sparsity and wireless-link superposition.
- Future directions: Future systems will involve heterogeneous devices cooperating to learn distinct but related functions in increasingly complex mobile environments.The choice between opportunistic and cooperative behavior depends largely on URLLC design.
- Future directions: FL is promising for cooperative automated driving through V2X interactions and collaborative robotics in complex, dynamic workflows.
APPENDIX: FEDERATED DATA AND ML MODEL
The appendix describes a radar-based collaborative-robotics setup in which agents classify human–robot situations from range and direction information. The decentralized FL model uses raw range-azimuth inputs, a compact two-layer network, and gossip or diffusion updates.
- Radar setup: The agents classify distance and direction-of-arrival information into six human–robot collaborative situations representing safe or unsafe conditions.
- Data: The simplified testing database is organized into four main structures and is available through the cited repository.
- Data: The testing database contains 900 range-azimuth measurements of size 256 × 63, with ranges from 0.5 to 11 m and angle bins from -75° to +75°.The corresponding test labels are integers from 0 to 5.
- Data: The training database also uses 900 true labels for its samples, with classes represented by integers from 0 to 5.
- ML model: The model takes background-subtracted 256 × 63 range-azimuth data and contains two trainable neural-network layers totaling 25,000 parameters.Decentralized FL uses gossip or diffusion with SGD step size µ_t = 0.025.