The smart factory has become one of the most durable images of contemporary industrial policy. Autonomous mobile robots circulate fluidly across production floors. Assembly lines dissolve into reconfigurable zones. Intelligence migrates outward, from centralized control rooms to the edge of the network and back again. Fifth‑generation cellular networks are repeatedly described as the enabling substrate that finally reconciles mobility with determinism, wireless flexibility with industrial control. Yet the distance between this vision and what current technology reliably delivers remains poorly understood in policy debate and industrial roadmaps.
This article examines that gap through the lens of recent research on Factory 5G, specifically the doctoral work of Nils Jörgensen at KTH Royal Institute of Technology. Rather than treating 5G as a faster or more reliable wireless pipe, the thesis interrogates what happens when mobile robots, industrial task planning, and cellular network resource management are treated as parts of a single cyber‑physical system. The result is neither a dismissal of 5G nor a repetition of vendor promises, but a careful diagnosis of where current research assumptions diverge from deployed reality, and why those divergences matter for industrial decision‑makers.
The argument that emerges is subtle but consequential. The principal challenge for Factory 5G is not radio propagation, latency budgets, or peak data rates in isolation. It is the lack of planning frameworks that operate at the same level of abstraction as real industrial work. Robots do not merely move, they execute missions. Networks do not merely transmit signals, they provide services with finite, schedulable resources. When these two facts are ignored, algorithmic sophistication accumulates on top of mistaken premises. When they are acknowledged, a different class of planning problem comes into view.
Industry 4.0 as a coordination problem
The popular framing of Industry 4.0 often emphasizes flexibility, customization, and digitalization. More analytically, it represents a shift from static production structures toward continuously reconfigurable socio‑technical systems. Traditional factories optimize for throughput under fixed workflows. Smart factories optimize for responsiveness under changing tasks. This shift alters the computational core of manufacturing.
Autonomous mobile robots exemplify this change. Unlike fixed industrial manipulators, automated guided vehicles and autonomous mobile robots make decisions about where to go, when to act, and how to coordinate with other machines. Global robot installations doubled between 2014 and 2024, but this growth has outpaced the maturation of fleet‑level coordination software. Surveys consistently show that navigation is largely solved, while integrated task assignment, scheduling, and collaboration remain open problems in practice.
This imbalance has architectural consequences. Coordination at scale rarely emerges from purely decentralized algorithms. Industrial deployments instead consolidate planning at shared compute layers, most often at the edge. Task allocation, replanning, and global state synchronization migrate upward, while individual robots execute local motion controllers. Edge computing thus becomes the cognitive centre of the factory.
Once that shift occurs, wireless communication ceases to be an auxiliary concern. It becomes load‑bearing infrastructure. Control messages, sensor streams, digital twin updates, and replanning triggers all traverse the same network fabric. The factory’s ability to coordinate itself becomes conditional on sustained, predictable connectivity.
Why 5G was expected to solve this
Fifth‑generation cellular technology entered this landscape with exceptionally high expectations. Unlike WiFi, 5G operates in licensed spectrum, supports fine‑grained scheduling, and introduces the concept of network slicing. Edge computing was designed into the core architecture rather than bolted on. Mobility management evolved from reactive handovers to proactive, make‑before‑break procedures. On paper, these features align closely with the needs of mobile industrial systems.
Industrial consortia such as 5G‑ACIA and standardization bodies within 3GPP codified these expectations early. White papers framed 5G as a deterministic substrate for automation. Vendors such as Ericsson, Nokia, and Huawei showcased private 5G deployments in controlled factory environments. Policymakers interpreted these demonstrations as evidence that wireless automation was no longer constrained by physics, only by adoption.
The thesis examined here does not dispute the technical merit of these features. Instead, it asks how they are actually used inside planning frameworks. It finds that most robotics and automation research do not interact with 5G at the level where its guarantees are implemented. This disconnection begins with abstraction.
The persistence of channel‑centric thinking
At the intersection of robotics and wireless communication, a substantial literature now exists under labels such as communication‑aware motion planning, motion‑communication co‑optimization, and joint trajectory‑communication optimization. Despite terminological diversity, a consistent pattern emerges. Robots plan their motion while optimizing or constraining a physical‑layer metric, most often signal‑to‑interference‑plus‑noise ratio or received power.
This approach is mathematically convenient. SINR varies smoothly with position, enabling gradient‑based optimization. It can be predicted using ray‑tracing or simplified path‑loss models. It fits naturally into trajectory planners and optimal control frameworks.
However, this convenience conceals a critical assumption. It presumes that favourable channel conditions correspond directly to favourable system‑level performance. In modern cellular networks, that presumption is increasingly false. Throughput, latency, and reliability depend not only on channel quality but on link adaptation, scheduling decisions, retransmissions, and spatial multiplexing behaviour that are invisible to channel‑centric models.
The doctoral work confronts this assumption empirically. In a private Ericsson 5G deployment, detailed ray‑tracing simulations predicted SINR with reasonable accuracy across an industrial floor. Yet measured throughput was consistently and significantly lower than predicted. The dominant source of error was not noise or interference but spatial multiplexing. The simulator assumed sustained four‑layer MIMO transmission. The real network frequently adapted down to one or two layers.
This discrepancy matters because throughput scales linearly with spatial rank. Even perfect SINR knowledge cannot compensate for optimistic multiplexing assumptions. The conclusion is not that simulation is useless, but that channel‑centric metrics cannot serve as reliable proxies for end‑to‑end service performance in industrial environments.
The mobile relay fallacy
An even deeper misalignment appears in work that treats robots as active elements of the network. Many planning frameworks assign robots roles as mobile relays or routers, repositioning them to maintain connectivity. This paradigm emerges naturally from ad‑hoc or mesh networking traditions.
In cellular systems, it is largely inapplicable. User equipment does not forward traffic. Scheduling, interference management, and resource allocation are centralized at the base station. While 5G includes sidelink mechanisms for device‑to‑device communication, industrial assessments by 5G‑ACIA conclude that these mechanisms have not been validated for ultra‑reliable low‑latency factory control.
The persistence of relay‑based planning reflects a deeper issue. Much of the joint planning literature abstracts away the network architecture it claims to address. Algorithms are developed against generic wireless graphs rather than against the resource model of cellular systems. The result is theoretical elegance divorced from deployable infrastructure.
From motion planning to mission planning
Perhaps the most consequential critique in the thesis concerns planning level, rather than networking. Industrial robots do not exist to move optimally through space. They exist to perform tasks under constraints. Transport material. Synchronize assembly steps. Meet deadlines. Collaborate in multi‑agent operations.
Yet most communication‑aware planners operate at the motion or trajectory level. They optimize where to go but not what to do or when to do it. Task allocation, sequencing, and temporal dependencies are either ignored or assumed externally solved.
The thesis proposes a conceptual shift. Rather than extending motion planners with communication metrics, extend mission planners with network service abstractions. In practical terms, this means adopting symbolic planning frameworks capable of representing tasks, resources, deadlines, and agent heterogeneity. The Planning Domain Definition Language is one such framework.
In this formulation, communication becomes a consumable resource alongside time, energy, and tool availability. Fifth‑generation physical resource blocks are explicitly represented as planner decision variables. Tasks consume spectrum for their duration. The planner searches over task assignments, robot paths, and network allocations simultaneously.
When evaluated in simulation, this joint formulation showed that mission‑level coordination can cut spectrum requirements by half while still meeting deadlines. More importantly, it demonstrated that network capacity need not be statically over‑provisioned. It can be planned.
Where assumptions meet infrastructure
The most revealing moment of the research was not an algorithmic improvement but an experimental failure. The joint planning framework assumed that radio‑level network slicing would be available in a state‑of‑the‑art private 5G system. In practice, it was not. Despite compliance with 3GPP standards, the deployed equipment did not support dynamic partitioning of physical resource blocks in the manner assumed by both the planner and much of the research literature.
This discovery reframes the problem yet again. Industrial planning frameworks often assume capabilities that exist in standards documents but not in shipping products. Policymakers and industrial strategists frequently conflate standardization with deployment. The thesis demonstrates that this conflation is consequential. Planning abstractions built atop absent features cannot be validated, regardless of their theoretical merits.
The implications are not pessimistic, but they are sobering. Private 5G remains a moving target. Standalone deployments with full slicing support are still rare. Lifecycle automation for slices is immature. Cross‑domain integration between radio access and edge computing remains largely manual.
Data‑driven alternatives and their limits
The measurement campaign did uncover one promising alternative. Machine‑learning models trained directly on throughput measurements were able to reduce prediction error substantially and eliminate systematic bias. These models bypassed channel‑centric assumptions entirely.
Yet this solution also carries trade‑offs. Gaussian process regression scales poorly with data volume. It adapts slowly to environmental change. A factory reconfiguration may invalidate learned models precisely in the regimes where planning robustness matters most.
The lesson is not that data‑driven models should replace physical insight, but that planning requires abstraction layers aligned with controllable interfaces. Learning end‑to‑end behaviour is a stopgap, not a substitute for service‑level guarantees.
Intent‑based networks and the future of factory planning
The thesis ultimately gestures toward an architectural resolution rather than a specific algorithm. Instead of planners predicting network performance, networks should expose what they can guarantee. Mission planners should express intent, not configuration. Sustain 50 megabits per second for this task. Maintain bounded latency for this collaborative operation.
This vision aligns with emerging work on intent‑based networking and autonomous network management. It also aligns with the direction of AI‑native radio access networks, where scheduling, beamforming, and resource allocation become adaptive services rather than static parameters.
If such architectures mature, joint communication and mission planning becomes less about optimization and more about negotiation. Planners operate within feasible envelopes returned by the network. Networks allocate resources in response to declared demands. The complexity boundary shifts to where it can be managed.
Policy and industrial implications
For policymakers, the central implication is that Factory 5G cannot be evaluated solely through coverage maps or latency targets. Its value depends on whether its capabilities can be referenced, planned, and verified at the level of industrial tasks. Investment in private networks must be matched with investment in integration frameworks.
For industry leaders, the lesson is more operational. Wireless automation succeeds when planning assumptions reflect deployed capabilities. Simulation alone is insufficient. Measurement‑grounded evaluation is not optional. Vendors, integrators, and operators must align on which features are standard, which are experimental, and which exist only on slide decks.
For researchers, the thesis offers a methodological caution. Algorithmic novelty does not compensate for abstraction mismatch. Motion‑level planning with channel‑level metrics will not scale to mission‑level industrial coordination, regardless of solver sophistication.
The challenge of Factory 5G is not that wireless networks are unreliable or robots are insufficiently autonomous. It is that planning frameworks have not caught up with the layered reality of industrial systems. Communication is a service. Tasks are missions. Guarantees emerge from scheduling, not from signal strength.
Recognizing this does not diminish the role of 5G. It clarifies it. The smart factory will not be built by faster radios alone, but by planning architectures that treat networks and robots as co‑dependents within the same operational logic. That shift in perspective may prove more decisive than any single technological breakthrough.
References
Jörgensen, N. (2026). Joint communication and mission planning: The real‑world challenges of Factory 5G. Doctoral dissertation, KTH Royal Institute of Technology.
Sánchez, J. M. G., Jörgensen, N., et al. (2022). Edge computing for cyber‑physical systems: A systematic mapping study emphasizing trustworthiness. ACM Transactions on Cyber‑Physical Systems, 6(3). https://doi.org/10.1145/3539662
Jörgensen, N., Kattepur, A., Mohalik, S., Vulgarakis, A., & Fersman, E. (2022). Towards 5G‑aware robot planning for industrial applications. Proceedings of the IEEE International Conference on Emerging Technologies and Factory Automation. https://doi.org/10.1109/ETFA52439.2022.9921449
Jörgensen, N., Kattepur, A., Mohalik, S., Vulgarakis, A., & Fersman, E. (2023). RoboPlan5G: Coordinating cloud‑controlled mobile robots with 5G network configuration. Proceedings of the IEEE International Conference on Emerging Technologies and Factory Automation. https://doi.org/10.1109/ETFA54631.2023.10275420
Jörgensen, N. (2026). Why channel‑centric models are not enough to predict end‑to‑end performance in private 5G: A measurement campaign and case study. arXiv preprint arXiv:2603.08865
Chen, K.‑C., Lin, S.‑C., Hsiao, J.‑H., Liu, C.‑H., Molisch, A. F., & Fettweis, G. P. (2021). Wireless networked multirobot systems in smart factories. Proceedings of the IEEE, 109(4), 468–494. https://doi.org/10.1109/JPROC.2020.3033753
Mahmood, A., Abedin, S. F., Sauter, T., Gidlund, M., & Landernäs, K. (2022). Factory 5G: A review of industry‑centric features and deployment options. IEEE Industrial Electronics Magazine, 16(2), 16–29. https://doi.org/10.1109/MIE.2022.3149209
Antonyshyn, L., Silveira, J., Givigi, S., & Marshall, J. (2023). Multiple mobile robot task and motion planning: A survey. ACM Computing Surveys, 55(10). https://doi.org/10.1145/3564696
