Federated Split Learning for Resource-Constrained Robots in Industrial IoT: Framework Comparison, Optimization Strategies, and Future Directions

TL;DR

Federated Split Learning (FedSL) enables resource-constrained industrial robots to collaboratively train large AI models while preserving data privacy and reducing on-device computation. The paper provides a comprehensive framework-by-framework comparison (synchronous, asynchronous, hierarchical, heterogeneous), a taxonomy of multimodal token fusion (pre-, intra-, post-fusion), and adaptive optimization techniques (model architecture, split-layer choice, computing frequency, and radio resource management). Simulation in a warehouse and a Carla-based indoor factory demonstrate performance trends, robustness to packet loss, and privacy-preserving decision-making with LLM assistance. The work identifies open challenges and directions for dynamic environments, privacy protections, human-in-the-loop collaboration, and energy-aware scheduling to advance sustainable, intelligent industrial systems.

Abstract

Federated split learning (FedSL) has emerged as a promising paradigm for enabling collaborative intelligence in industrial Internet of Things (IoT) systems, particularly in smart factories where data privacy, communication efficiency, and device heterogeneity are critical concerns. In this article, we present a comprehensive study of FedSL frameworks tailored for resource-constrained robots in industrial scenarios. We compare synchronous, asynchronous, hierarchical, and heterogeneous FedSL frameworks in terms of workflow, scalability, adaptability, and limitations under dynamic industrial conditions. Furthermore, we systematically categorize token fusion strategies into three paradigms: input-level (pre-fusion), intermediate-level (intra-fusion), and output-level (post-fusion), and summarize their respective strengths in industrial applications. We also provide adaptive optimization techniques to enhance the efficiency and feasibility of FedSL implementation, including model compression, split layer selection, computing frequency allocation, and wireless resource management. Simulation results validate the performance of these frameworks under industrial detection scenarios. Finally, we outline open issues and research directions of FedSL in future smart manufacturing systems.

Figures (5)

• Figure 1: (a) An illustration of split lerning. (b) The implementation of vanilla FedSL in industrial environments faces several key technical challenges, including data heterogeneity, device heterogeneity, model heterogeneity, and communication heterogeneity. i) Data heterogeneity arises from the imbalance data across devices, where local datasets may exhibit significant statistical divergence due to varying usage patterns or environments. ii) Device heterogeneity refers to the disparities in computational power, memory capacity, and energy constraints among robots. iii) Model heterogeneity occurs when personalized AI models with different architectures or parameter dimensions are deployed across devices. iv) Lastly, communication heterogeneity stems from the variability in network connectivity, such as 4G, 5G, and Wi-Fi, resulting in fluctuating bandwidth, latency, and reliability during model updates and intermediate data transmission. (c) FedSL for resource-constrained robots in industrial IoT systems within a three-tier robot-edge-cloud architecture. In this article, we use the terms "device", "client" and "robot" interchangeably, which does not affect the readability.
• Figure 2: Framework comparison of different FedSL paradigms. (a) An illustration of the workflow of synchronous or asynchronous FedSL. (b) Hierarchical FedSL, which decomposes large-scale AI models into three tiers. End devices are responsible for lightweight feature extraction, edge servers handle intermediate computational tasks, and the cloud performs high-level, complex reasoning processes. (c) Heterogeneous FedSL, which addresses model heterogeneity across IIoT clusters (e.g., different factories or product lines). Within each cluster, a specialized model is trained using the standard FedSL approach, while cross-cluster knowledge integration is achieved through cloud-based distillation and sharing mechanisms.
• Figure 3: Three token fusion strategies: (a) input-level token fusion, (b) intermediate-level token fusion, and (c) output-level token fusion.
• Figure 4: Training performance of FedSL frameworks in industrial systems.
• Figure 5: Perception, decision, and control profiles of different schemes in a Carla-based indoor factory simulation. (a) Scenario setup with an ego-robot navigating a corridor with 4 obstacles. (b) FedSL enables accurate obstacle detection and safe right detour via federated models and LLM-based decision-making. (c) Pure federated learning without high-level LLM reasoning. (d) Pure split learning with pretrained model.