Paper deep dive
World-Model-Aware Responsibility Allocation in Heterogeneous Logistics Systems
Artan Markaj, Niklas Jobs, Felix Gehlhoff
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 91%
Last extracted: 7/17/2026, 2:09:43 AM
Summary
The paper introduces the World-Model-Aware Responsibility Framework (WMARF) to dynamically allocate decision authority between autonomous logistic equipment (ALE) and a central control system (CS) in heterogeneous logistics environments. By evaluating world-model quality and automation levels, WMARF prevents deadlocks caused by divergent beliefs and classifies them by authority state. Simulations and VDA 5050 interface tests demonstrate that proximity-triggered handoffs effectively resolve governance failures without requiring static protocol agreements.
Entities (10)
Relation Signals (10)
WMARF â assigns â decision authority
confidence 95% ¡ WMARF, which assigns authority dynamically from CS world-model quality and equipment automation level
WMARF â prevents â deadlocks
confidence 93% ¡ a divergence deadlock under static control is prevented by a proximity-triggered handoff
WMARF â classifies â deadlocks
confidence 92% ¡ classifies deadlocks by the state of authority -- none, in transition, or divergent
WMARF â augments â Central Control System (CS)
confidence 91% ¡ WMARF augments, rather than replaces, an existing CS: a thin mediator runs alongside the CS
Central Control System (CS) â maintains â global world model
confidence 90% ¡ The CS maintains a global world model and issues coordination decisions
Quadrant Classifier â determines â authority zones
confidence 89% ¡ The two axes yield four authority zones. In Q1... Q2... Q3... Q4...
Autonomous Logistic Equipment (ALE) â operatesalongside â non-autonomous machinery
confidence 88% ¡ Logistics systems increasingly mix autonomous logistic equipment (ALE) with non-autonomous machinery
Handoff Evaluator â triggers â authority transfer
confidence 87% ¡ the Handoff Evaluator transfers authority to it, its sensor reading reconciles the CS model
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Logistics systems increasingly mix \emph{autonomous logistic equipment} (ALE) with non-autonomous machinery under a central control system (CS), where the best decision-maker depends on who holds the most current world model, yet authority is fixed at design time. When an ALE's local model and the CS global model diverge, both act on incompatible beliefs and produce deadlocks that resource-based handling neither explains nor prevents. We propose the World-Model-Aware Responsibility Framework (WMARF), which assigns authority dynamically from CS world-model quality and equipment automation level, and classifies deadlocks by the state of authority -- none, in transition, or divergent. In a discrete-event simulation of two ALE converging on a semi-automated transfer point, reproduced over the VDA~5050 interface, a divergence deadlock under static control is prevented by a proximity-triggered handoff. Because authority follows information quality rather than a shared protocol, the scheme stays valid as autonomy grows.
Tags
Links
- Source: https://arxiv.org/abs/2607.14550v1
- Canonical: https://arxiv.org/abs/2607.14550v1
Trouble viewing inline? Open PDF directly â
Full Text
21,945 characters extracted from source content.
Expand or collapse full text
World-Model-Aware Responsibility Allocation in Heterogeneous Logistics Systems Artan Markaj Eurogate GmbH & Co. KGaA, KG Hamburg, Germany artan.markaj@eurogate.eu Niklas Jobs Helmut Schmidt University Hamburg Hamburg, Germany niklas.jobs@hsu-h.de Felix Gehlhoff Helmut Schmidt University Hamburg Hamburg, Germany felix.gehlhoff@hsu-h.de AbstractâLogistics systems increasingly mix autonomous lo- gistic equipment (ALE) with non-autonomous machinery under a central control system (CS), where the best decision-maker depends on who holds the most current world model, yet authority is fixed at design time. When an ALEâs local model and the CS global model diverge, both act on incompatible beliefs and produce deadlocks that resource-based handling neither explains nor prevents. We propose the World-Model-Aware Responsibility Framework (WMARF), which assigns authority dynamically from CS world-model quality and equipment automation level, and classifies deadlocks by the state of authority â none, in transition, or divergent. In a discrete-event simulation of two ALE converging on a semi-automated transfer point, reproduced over the VDA 5050 interface, a divergence deadlock under static control is prevented by a proximity-triggered handoff. Because authority follows information quality rather than a shared protocol, the scheme stays valid as autonomy grows. Index Termsâautonomous logistic equipment, intralogistics, heterogeneous automation, decision authority, world model, dead- lock, agent-based control I. INTRODUCTION Logistics systems increasingly combine equipment at very different levels of automation. Autonomous logistic equip- ment (ALE) (e.g., automated guided vehicles, autonomous mobile robots, and increasingly autonomous trucks) operates alongside non- or semi-autonomous equipment such as con- veyors, handling stations, and manually operated machinery, all coordinated by a central control system (CS) [1], [2]. The CS maintains a global world model and issues coor- dination decisions, while autonomous equipment carries a rich, real-time local world model from onboard sensors; non- autonomous equipment, by contrast, contributes only coarse state information whose granularity falls with its automation level. This heterogeneity creates a structural problem that current control architectures do not address: which entity should hold decision authority is not fixed, it depends on who has the most complete and current world model for the decision at hand. Yet responsibility is almost always allocated statically, at design time, from the automation landscape as it stands then [3], Š 2026 IEEE. Personal use of this material is permitted. Permission from IEEE must be obtained for all other uses, in any current or future media, including reprinting/republishing this material for advertising or promotional purposes, creating new collective works, for resale or redistribution to servers or lists, or reuse of any copyrighted component of this work in other works. not from the information quality actually available at runtime. When the local model of an ALE diverges from the CS model (through latency, sensor-coverage gaps, or equipment that does not report fine-grained state), neither entity has clear authority, and both may act on locally consistent but globally incompatible beliefs [4], proceeding at once into a deadlock that no party flagged. Two forces make this acute. First, a maturity trap: au- tonomous equipment will grow markedly more capable within a systemâs operational life, so a centralized architecture chosen today for safety can become tomorrowâs bottleneck, while a decentralized one chosen for future scalability may be unsafe now (current ALE may lack the sensing or reasoning maturity to act safely without CS oversight). Second, a multi- vendor reality: equipment from different manufacturers ex- poses incompatible interfaces and state models, so neither full centralization (one common protocol) nor full decentralization (every device negotiating autonomously) is achievable. We address this with the World-Model-Aware Responsi- bility Framework (WMARF), a reference architecture for runtime authority assignment. It raises two questions: whether decision authority between autonomous equipment and the CS can be assigned dynamically from world-model quality and automation level (RQ1), and whether the resulting deadlocks can be classified by the state of authority so that each admits a targeted resolution (RQ2). Accordingly we make two contribu- tions: (i) WMARF, a world-model-aware scheme that assigns authority dynamically and labels the resulting deadlocks by their authority state â none, in transition, or divergent; and (i) the framing of the maturity trap as a design problem, with a migration path that keeps the allocation valid as equipment autonomy grows without architectural redesign, and without requiring vendors to agree on a protocol. I. BACKGROUND AND RELATED WORK Coordination of equipment in logistics has matured along two poles. At one, a central control system (CS) schedules, routes, and dispatches from a global model [1], [3], in the tradition of human supervisory control [5]; at the other, ALE negotiates locally and decentralizes decisions to react to a changing environment [2], [6]. Decentralized, agent-based schemes increasingly reason about imperfect information [4], yet still assume the deciding entities are capable agents. arXiv:2607.14550v1 [cs.MA] 16 Jul 2026 Reassigning authority to whoever is best placed to decide is the subject of variable autonomy [7] â a topic spanning adjustable and sliding autonomy [8], mixed-initiative interac- tion [9], and dynamic role and task allocation for humanârobot teams [10], [11], grounded in classic models of automation levels and function allocation [12], and recently extended to learn when to adjust the level of automation online [13]. This body almost exclusively arbitrates between a human and an autonomous system, keying the hand-over on human workload, trust, or task demand. Two of its assumptions break in our setting: authority shifts between software systems (the CS and the ALE), not between a human and a software system; and one party â non-autonomous equipment that only reports coarse state â is not a capable agent at all. Deadlocks, in turn, are treated as resource or routing conflicts, from classical AGV conflict resolution [14], [15] to recent learning- and barrier- based avoidance [16], [17], and are never classified by who should have been deciding. This leaves a gap that real installations occupy. Much work naturally concentrates on the two autonomous extremes, fully centralized or fully decentralized, while many deployed systems operate in the middle ground: mixed fleets from sev- eral vendors, ALE sharing space with conveyors, cranes, and manual stations whose state the CS barely observes. Heteroge- neous multi-robot systems are themselves well surveyed [18], but they coordinate robots; here the heterogeneity spans autonomous and non-autonomous equipment under one CS. Three failure patterns characterize it. A governance vacuum arises when local and global models diverge and no protocol says who resolves it. Automation-level blindness arises when the CS represents all equipment uniformly, so decisions near poorly observable equipment are systematically overconfident. And many deadlocks are governance failures: two entities act at once because authority was never transferred, not because a resource is busy. WMARF targets this heterogeneous middle ground. I. THE WMARF FRAMEWORK WMARF assigns decision authority dynamically by posi- tioning each interaction between an ALE and a target equip- ment in a two-dimensional space (Fig. 1). The structural axis is the targetâs automation level and, with it, how observable its state is to the CS through its interface. Automation and observability usually rise together, but not always: a highly automated device from another vendor may still expose only coarse state and therefore sit lower on this axis than its au- tonomy alone would suggest. The dynamic axis is the current quality â recency and consistency â of the CS world model for that target. Reading the structural axis as observability is what lets WMARF absorb the multi-vendor problem: authority follows how well a target can be observed, not which vendor built it or whether a common protocol exists. A. Responsibility quadrants The two axes yield four authority zones. In Q1 (high quality, high automation) the CS model is reliable and the ALE follows Q2 â Shared / Handoff both valid but may diverge; explicit protocol Q1 â CS-dominant CS model reliable; ALE follows CS Q4 â Contested no reliable model; halt + escalate Q3 â ALE-dominant onboard sensors are ground truth CS world-model quality High Low Equipment automation level Low High Fig. 1. The WMARF responsibility quadrants. Authority follows the entity with the better world model; an interactionâs position shifts at runtime with proximity, model staleness, and sensor-vs-CS divergence. TABLE I EXEMPLARY HANDOFF TRIGGERS (NON-EXHAUSTIVE). TriggerDirectionCondition ProximityCSâALEdistance < θ DivergenceCSâALE |sensorâ CS| > δ Automation levelCSâALEtarget manual/semi StalenessCSâALECS update age > Ď Resolution completeALEâCSinteraction finished EscalationALEâCSALE detects deadlock CS instructions. In Q2 (high quality, low automation) the CS has a good global picture but the targetâs internal state is under- represented; a handoff (transfer of decision authority) gives the ALE local authority for the final approach, after which it reports back and the CS updates its model. In Q3 (low quality, high automation) the CS model is stale or inconsistent â post- fault, during reconfiguration, or when update latency exceeds an age Ď â and the ALE trusts its onboard sensors under a mandatory escalation flag. In Q4 (low quality, low automation) neither entity has a reliable model and the conservative default â halt, escalate, await clearance â applies. Position is not static: it shifts at runtime with ALE proximity, model staleness, and the discrepancy between sensor readings and CS belief. B. Handoff triggers Authority transfers fire on measurable conditions. Table I lists an exemplary, non-exhaustive set: CSâALE triggers on proximity, model divergence, target automation class, or CS staleness, and ALEâCS triggers on completion or a locally detected deadlock; each transfer is closed by an acknowledg- ment handshake. Identifying the full trigger set and calibrating its thresholds is where the principal work of this framework lies, and is the focus of future work. C. Deadlock taxonomy by authority state We classify deadlocks by the state of decision authority rather than the contested resource. Type A (no authority as- signed): both entities act on their own beliefs with no handoff TABLE I DEADLOCK TAXONOMY BY AUTHORITY STATE. TypeAuthorityArises inResolution Anone assignedQ4halt + escalate Bin transitionany handoffacknowledgment handshake Cdivergent modelsQ3 / Q2modelreconciliation (sensorprecedence within θ) engaged, a governance vacuum, typical of Q4; resolved by the conservative default firing before either commits. Type B (authority in transition): authority is mid-transfer and the receiver has not yet acknowledged or updated its model (e.g., the CS has issued a handoff token but the ALE has not yet committed to the new authority), leaving a decision vacuum in the transfer window; resolved by the acknowledgment hand- shake. Type C (authority on divergent models): authority is held, but the holders act on locally consistent yet globally incompatible models; resolved by model reconciliation, with sensor data taking precedence within proximity θ of the target. The quadrants indicate where each type tends to arise, not its definition (Table I). D. Maturity migration As equipment matures, the thresholds θ,δ,Ď can be retuned to enlarge the ALE-dominant zone without changing the architecture; in the limits WMARF degrades gracefully toward full centralization (an omniscient CS) or full decentralization (fully autonomous equipment). This is a possible answer to the maturity trap, and, because authority follows information quality rather than a shared protocol, to the multi-vendor problem. IV. ARCHITECTURE AND IMPLEMENTATION WMARF augments, rather than replaces, an existing CS: a thin mediator runs alongside the CS, with a lightweight client on each ALE (Fig. 2). It reads the CS world-model entry for a target together with its timestamp, and the ALEâs proximity and onboard sensor reading; it emits an authority token â CS- holds or ALE-holds â for that interaction, and, on trouble, a deadlock label and a resolution. The CS keeps dispatching and the ALE keeps acting (e.g., driving); WMARF only decides whose decision counts, and pushes a corrected state back when the ALEâs sensors win. Integration uses the fleetâs existing interface (e.g., VDA 5050), so no control logic is rewritten and no shared protocol is imposed. A. Implementation and scenario We implemented the three components of Fig. 2 as Python modules (a Quadrant Classifier, a Handoff Trigger Evaluator, and a Deadlock Classifier) driven by a discrete-event simula- tion in SimPy. The scenario is the canonical hard case: the CS dispatches two ALE to the same semi-automated transfer point, which it believes is ready from a stale update, while the point is in fact still in a mechanical transition that its interface Control System (CS) [existing] global world model ¡ dispatch / routing Handoff Evaluator Quadrant Classifier Deadlock Classifier WMARF Mediator (augments, not replaces) ALE [existing] onboard sensors ¡ local model Target equipment/zone conveyor / crane / transfer point state + recency reroute / hold telemetry, sensor authority token interact Fig. 2. WMARF deployment. The mediator augments the existing CS and decides, per interaction, whether the CS or the ALE holds authority; it never moves equipment or rewrites control logic. ALE-2 ALE-1 DEADLOCK (Type C) (a) Without WMARF MOVINGWAITHOLDUSINGBLOCKED 051015202530 time (s) ALE-2 ALE-1 handoff -> reconcile -> reroute (b) With WMARF Fig. 3. The transfer-point scenario, the same world run twice. (a) Without WMARF, both ALE commit on the stale model and a Type-C divergence deadlock occurs at t = 12 s. (b) With WMARF, a proximity handoff at t = 7 s reconciles the CS model and holds the second ALE, and both transfers complete. All values are produced by running the simulation. does not expose. WMARF classifies the situation as Q2, a good global picture but an under-represented internal state, where the handoff protocol should engage. B. Result Figure 3 reports both runs of the same world. Without WMARF, both ALE commit on the stale âreadyâ belief; the second commits at t = 12 s while the point is not serviceable, and the Deadlock Classifier labels a Type-C deadlock (au- thority held on divergent models), from which neither transfer completes. With WMARF, the first ALE crosses the proximity threshold at t = 7 s, the Handoff Evaluator transfers authority to it, its sensor reading reconciles the CS model, and the CS holds the second ALE; both transfers then complete and no deadlock occurs. The same trigger logic that surfaces the diver- gence also prevents it â evidence for RQ1 (dynamic authority averts the failure) and RQ2 (the deadlock is classifiable by authority state). This is a single, qualitative demonstration; a quantitative study across scenarios is future work. CS WMARF ALE-1ALE-2 order order statepoint: transition [decide: authority â ALE-1] reconcile (CS belief) instantAction startPause state: transfer complete instantAction stopPause state: transfer complete Fig. 4.VDA 5050 message exchange for the WMARF reproduction: the CS issues orders; ALE-1âs state reveals the divergence; WMARF takes the handoff, reconciles the CS belief, and sequences ALE-2 with startPause/stopPause, all over the standard interface, the WMARF core unchanged. C. Reproduction on the VDA 5050 interface To confirm that WMARF attaches to a real industrial interface and not only to the simulator, we re-ran the same scenario over VDA 5050 [19], the standard interface between a master control and AGVs (Fig. 4). Only the input/output adapter changes: the Quadrant Classifier, Handoff Evaluator, and Deadlock Classifier are reused unchanged, so WMARF augments the existing interface rather than replacing control logic. The ALE state heartbeat is the recency signal; the bracketed [decide:] annotation in Fig. 4 marks WMARFâs internal authority decision, which triggers the subsequent handoff messages that reconcile the CS belief and sequence the second ALE with instantActions. The run reproduces Fig. 3, the same Q2 handoff prevents the Type-C deadlock, over an in-process broker that can be swapped for MQTT. V. DISCUSSION AND OUTLOOK WMARF reframes a class of logistics deadlocks as gov- ernance failures and shows, on one scenario, that assigning authority by world-model quality averts a divergence deadlock that static control does not. The evidence is deliberately limited: a single qualitative scenario with two ALE and one target, a minimal deadlock model, and a mediator assumed reliable and above the safety-critical path. Above all, the trigger thresholds θ, δ, and Ď are set by hand; identifying the full trigger set and calibrating these thresholds is the principal open problem. The trigger taxonomy of Table I is domain-agnostic; its thresholds are the adaptation point for each deployment. Three lines of work follow: a quantitative study of WMARF against static centralized and decentralized baselines â dead- lock rate, throughput, and resolution time across heterogeneity levels; a formal treatment of the handoff protocol, covering convergence, deadlock-freedom, and principled threshold cal- ibration; and a testbed â a fleet manager over VDA 5050 [19] and Open-RMF [20], whose resource locking WMARF com- plements with an authority layer â scaled to many ALE and targets and richer non-autonomous-equipment state. Because the same thresholds can be retuned to widen the ALE- dominant zone as equipment matures, WMARF offers an allocation that stays valid as autonomy grows, a key practical advantage. REFERENCES [1] G. Fragapane, R. de Koster, F. Sgarbossa, and J. O. Strandhagen, âPlanning and control of autonomous mobile robots for intralogistics: Literature review and research agenda,â European Journal of Opera- tional Research, vol. 294, no. 2, p. 405â426, 2021. [2] R. Keith and H. M. La, âReview of autonomous mobile robots for the warehouse environment,â arXiv preprint arXiv:2406.08333, 2024. [3] F. Gehlhoff, N. Jobs, and V. Henkel, âAgent-based control of interac- tion areas in intralogistics: Concept, implementation and simulation,â Logistics, vol. 9, no. 2, 2025. [4] F. Gehlhoff, T. Busert, M. Lewin, and A. Fay, âIncorporating imper- fect information in decentralized agent-based dynamic and integrated scheduling,â in 2018 IEEE 23rd International Conference on Emerging Technologies and Factory Automation (ETFA), Torino, Italy, 2018, p. 816â823. [5] T. B. Sheridan, Telerobotics, Automation, and Human Supervisory Control. MIT Press, 1992. [6] A. Meseguer Valenzuela and F. Blanes Noguera, âTask allocation in mobile robot fleets: A review,â arXiv preprint arXiv:2501.08726, 2025. [7] A. Theodorou, M. Chiou, B. Lacerda, and S. RothfuĂ, âEditorial: Variable autonomy for human-robot teaming,â Frontiers in Robotics and AI, vol. 11, p. 1465183, 2024. [8] B. Sellner, F. W. Heger, L. M. Hiatt, R. Simmons, and S. Singh, âCoordinated multi-agent teams and sliding autonomy for large-scale assembly,â Proceedings of the IEEE, vol. 94, no. 7, p. 1425â1444, 2006. [9] E. Horvitz, âMixed-initiative interaction,â IEEE Intelligent Systems, vol. 14, no. 5, p. 14â24, 1999. [10] E. Lamon, F. Fusaro, E. De Momi, and A. Ajoudani, âA unified architecture for dynamic role allocation and collaborative task planning in mixed human-robot teams,â arXiv preprint arXiv:2301.08038, 2023. [11] Z. Yuan, R. Wang, T. Kim, D. Zhao, I. Obi, and B.-C. Min, âAdaptive task allocation in multi-human multi-robot teams under team heterogene- ity and dynamic information uncertainty,â in 2025 IEEE International Conference on Robotics and Automation (ICRA), 2025. [12] R. Parasuraman, T. B. Sheridan, and C. D. Wickens, âA model for types and levels of human interaction with automation,â IEEE Transactions on Systems, Man, and CyberneticsâPart A: Systems and Humans, vol. 30, no. 3, p. 286â297, 2000. [13] M. Hajnorouzi, A. Rakow, and M. Fr Ě anzle, âModel learning for adjusting the level of automation in HCPS,â in Formal Methods for Autonomous Systems (FMAS), ser. EPTCS, vol. 436, 2025. [14] I. F. A. Vis, âSurvey of research in the design and control of automated guided vehicle systems,â European Journal of Operational Research, vol. 170, no. 3, p. 677â709, 2006. [15] S. A. Reveliotis, âConflict resolution in AGV systems,â IIE Transactions, vol. 32, no. 7, p. 647â659, 2000. [16] M. M Ě uller, âMulti-agent reinforcement learning for deadlock handling among autonomous mobile robots,â arXiv preprint arXiv:2511.07071, 2025. [17] Y. Zhang, Y. Lyu, S. Jo, Y. Yang, and W. Luo, âAdaptive deadlock avoidance for decentralized multi-agent systems via CBF-inspired risk measurement,â in 2025 IEEE International Conference on Robotics and Automation (ICRA), 2025. [18] Y. Rizk, M. Awad, and E. W. Tunstel, âCooperative heterogeneous multi- robot systems: A survey,â ACM Computing Surveys, vol. 52, no. 2, p. 1â31, 2019. [19] German Association of the Automotive Industry (VDA), âVDA 5050: Interface for the communication between automated guided vehicles (AGV) and a master control,â Standard, 2022, version 2.0. [20] Open Source Robotics Foundation, âOpen-RMF: Open robotics middle- ware framework,â Open-source project, 2024.