DAG-Based QoS-Aware Dynamic Task Placement for Networked Multi-Stage Control Pipelines

DAG-Based QoS-Aware Dynamic Task Placement for Networked Multi-Stage Control Pipelines Thien Tran∗ , Jonathan Kua∗ , Thuong Hoang∗ , Minh Tran† , Yuemin Ding‡ , and Jiong Jin§ ∗ Deakin University, Australia; † RMIT University, Vietnam; ‡ University of Navarra, Spain; § Swinburne University of Technology, Australia

arXiv:2605.19887v1 [cs.DC] 19 May 2026

{peter.tran, jonathan.kua, thuong.hoang}@deakin.edu.au; [email protected]; [email protected]; [email protected]

Abstract—Current Physical AI (PAI) relies heavily on closedloop visual-servoing pipelines, whose perception and planning stages may become computationally intensive onboard due to complex models embedded on robots. In practice, offloading the perception task to on-site edges statically is inappropriate for latency-sensitive, precise industrial settings over a standardized industrial network. This emphasizes the importance of Control– Communication–Computing (3C) co-design in industrial automation: monolithic local execution saturates AI-accelerated machine and robot hardware, while static edge offloading exposes the control loop to network jitter. Existing adaptive task placement (ATP) controllers can partially address the gap by relocating a single pipeline stage on binary threshold rules, without a multi-stage model and an explicit cost on placement switching. In this Work-in-Progress (WiP) paper, we propose a directed acyclic graph (DAG) based quality-of-service (QoS)-aware dynamic task placement (DTP) framework for sensing–perception– planning–control pipelines in networked robotics. This pipeline is formalized as a DAG with task-level and node-level attributes for compute cost, communication delay, and feasible placement sets; over a small interpretable candidate set (fully local, static offload, hybrid), a window-based cost function combines tail endto-end latency, deadline violation rate, hardware utilization, and a Hamming-distance switching penalty, and a DTP algorithm with hysteresis and a minimum dwell-time bounds placement chatter. Our WiP paper presents the theoretical framework, a structured qualitative analysis, and a two-phase simulation-plus-hardwarein-the-loop validation roadmap. Index Terms—Directed Acyclic Graphs (DAG), QoS-Aware Dynamic Task Placement (DTP), Networked Robotics Systems

I. I NTRODUCTION The compute envelope of factory robotics is being reshaped by Physical AI (PAI) workloads. World Foundation Models (WFMs) for surrounding cognitive and agentic visual servoing demand inference costs that exceed the on-board envelope of typical industrial robots, while LLM-based task dispatchers exhibit cold-start latencies of sub-20 seconds, orders of magnitude beyond the sub-millisecond deadlines typical of industrial contact-rich control loops within Vision-LanguageAction (VLA) models embedded on PAI [1], [2]. This mismatch demands a unified networked orchestration framework that binds heterogeneous AI capabilities to the robot–edge compute fabric within the closed-loop timing budget. The Cloud-Fog Automation (CFA) reference architecture [3], [4] articulates a unified vision for networked industrial collaboration across the cloud–fog–edge continuum, while

deterministic substrates such as Time-Sensitive Networking (TSN) and 5G-URLLC [5] provide the underlying real-time fabric. Together, they make co-located edge servers a viable extension of the PAI compute envelope, provided placement decisions respect the closed-loop timing budget. However, the layer specifies the runtime decision rule: under workload and network variability that no design-time profile can predict, the deployment question shifts from “whether” to offload to “which” pipeline stages run, where by being decided online. A single static placement cannot serve this regime: both perception/planning compute demand and link delay are nonstationary, varying with scene complexity and model branching, and the cost of a wrong decision is asymmetric; a missed deadline propagates into actuation jitter that can violate safety envelopes. Empirical studies on industrial testbeds report deadline-violation rates above 40% under monolithic local deployment of visual-servoing workloads [6]. Adaptive placement is, therefore, a deployment necessity; the open question is what such a rule should observe what it should optimize, and how it should be stabilized for industrial deployment. DAG-based task offloading has been studied extensively in mobile edge computing (MEC) [7], [8], but these models target “batch workflows” and optimize average latency or energy rather than deadline violation rate (VD ) under closed-loop control. Our previous Adaptive task placement (ATP) on physical robot testbeds demonstrates the value of dynamic offloading, but each system relocates only a “single” pipeline stage using binary threshold rules on individual fixed metrics, providing no formal DAG model, no multi-objective cost function, and no analytical resistance to task placement oscillation. In this Work-in-Progress (WiP) paper, we propose a DAGbased QoS-aware DTP framework, the formal multi-objective extension of the empirically validated edge-based QoS-aware ATP controller from previous work. The system architectural framework is shown in Fig. 1. We present the complete theoretical framework, DAG model, cost function, and DTP algorithm, with the structured qualitative analysis and the validation roadmap. This work contributes the DAG pipeline model that distinguishes hard-anchored sensing and control stages from the relocatable perception and planning adaptation space, agnostic to the LAN/TSN/5G-URLLC substrate. On this model, we define the multi-objective QoS cost function and the window-based DTP algorithm with hysteresis and min-

(a) Pipeline DAG: Hard-Anchored Stages & Relocatable Adaptation Space

T1 Sensing

T2 Perception

T3 Planning

T4 Control

Anchored NT1 = {R1 }

Relocatable NT2 = {R1 , E}

Relocatable NT3 = {R2 , E}

Anchored NT4 = {R2 }

(b) Candidate Placements Πcand on the Compute Fabric

SO static offload

HYB

E Edge Server

R2 Heavy Robot

T1 , T2

–

T3 , T4

T1

T2 , T3

T4

T1

T2

T3 , T4

O BSERVE (window k, W cycles) {L95,k , VD,k , Ur,k , Ue,k } C OMPUTE COST Jk = αL L̃95 + αV VD + αrŨr + αeŨe + αS S D ECIDE & APPLY πk∗ = arg min Jk migrate iff ∆J > ∆min and nd ≥ Nmin

hybrid

apply πk∗ for W cycles

LOC fully local

R1 Light Robot

(c) Window-Based DTP Control Loop

Industrial real-time network: LAN / TSN / 5G-URLLC

Hard-Anchored (T1 , T4 )

Relocatable Adaptation Space (T2 , T3 )

DTP Control Loop on E

Fig. 1. DAG-QoS DTP system architecture. (a) The four-stage sensing–perception–planning–control pipeline as a DAG: T1 , T4 are hard-anchored; T2 , T3 form the relocatable adaptation space. (b) Three candidate placements (LOC, SO, HYB) on the R1 /E/R2 compute fabric. (c) Window-based DTP loop on E: per-window QoS observation, cost Jk , and decision πk∗ with hysteresis ∆min and dwell-time Nmin , maintaining VD ≤ 5%.

imum dwell-time that bounds placement chatter, formalizing the 3C principles within the single per-window optimization.

B. Placement Mapping and Architectural Invariant A placement mapping π : V → N assigns each task to a feasible node. The visual servoing pipeline imposes:

II. T HEORETICAL S YSTEM AND DAG- BASED M ODEL

NT1 = {R1 },

A. Compute Nodes and Workload DAG Let N = R ∪ E ∪ C denote the global compute node set, comprising robots – R, edge servers – E, and (optionally or extensively) cloud nodes – C. The target empirical factory configuration is R1 (light robot with monocular camera), R2 (heavy robot driving a manipulator), and E (edge server), interconnected by an industrial real-time network with configurable delay and jitter. The framework is agnostic to the link technology, engineered LAN, TSN, or 5G-URLLC [5], and is parameterized by the per-edge delay distribution dij (·). A control pipeline is modeled as a directed acyclic graph – G = (V, EG ), with task set V = {T1 , T2 , T3 , T4 } corresponding to four stages, and linear precedence EG = {(T1 , T2 ), (T2 , T3 ), (T3 , T4 )}. Each task v ∈ V carries: The feasible placement set Nv ⊆ N ; The compute-time function cv : Nv → R>0 ; and • The CPU-utilization contribution uv : Nv → [0, 1].

•

•

Denote that each precedence edge (vi , vj ) ∈ EG incurs a communication delay dij (ni , nj ) ≥ 0 depending on link characteristics and payload size of the associated operations.

NT2 = {R1 , E},

NT4 = {R2 },

NT3 = {R2 , E}.

(1) (2)

This encodes the key architectural invariant of the multi-PAI System in agentic industrial automation operation: Hard-anchored stages (T1 , T4 ) remain on physical AI-accelerated machines and robots, preserving submillisecond local operations regardless of network state. • Relocatable adaptation space (T2 , T3 ) migrates across the Cloud–Edge–Robotics continuum (native edge tiers) subject to the QoS cost function defined in Section III.

•

C. Control Loop and End-to-End (E2E) Latency The control loop is periodic with period P and deadline D ≤ P . Under placement π, the nominal E2E latency is: X X L(π) ≈ cv (π(v)) + dij (π(vi ), π(vj )). (3) v∈V

(vi ,vj )∈EG

Denote that L is stochastic in practice due to workload variability and network jitter during operations; high-percentile values therefore serve as the primary QoS-aware target.

III. Q O S-AWARE C OST M ODEL A. Per-Window QoS Metrics The native DTP algorithm operates over observation windows of W control cycles indexed by k. Per single window, the original four QoS metrics are collected as follows: • L95,k : Empirical 95th-percentile E2E latency, capturing tail behavior from network outliers and bursty compute; • VD,k ∈ [0, 1]: Deadline violation rate (primary SLA); • Ur,k , Ue,k : CPU utilization across robot/edge nodes. Latency and utilization are normalized by design targets: L95,k Ur,k Ue,k L̃95,k = , Ũr,k = , Ũe,k = . (4) L̄ Ūr Ūe VD,k is bounded in [0, 1] and requires no normalization. B. Switching Penalty and Cost Function To penalize frequent migration, a switching penalty based on the Hamming distance between placements is applied: 1 X  S(πk , πk−1 ) = 1 πk (v) ̸= πk−1 (v) . (5) |V| v∈V

This term suppresses “placement chatter”, high-frequency oscillation that degrades network stability and is the infrastructure-level manifestation of trust latency [9]. Given non-negative weights, the QoS-aware cost for window k is:

Jk = αL L̃95,k +αV VD,k +αr Ũr,k +αe Ũe,k +αS S(πk , πk−1 ). (6) The tension between αV VD,k (deadline aggression) and αS S(πk , πk−1 ) (migration resistance) formalizes the core control–infrastructure trade-off. In industrial settings, αV > αL > αS reflects prioritizing SLA compliance as the primary, raw latency as the secondary, and stability as a soft constraint. C. Constrained Placement Problem To keep per-window evaluation tractable on edge hardware, we originally restrict to a small interpretable candidate set Πcand comprising the following three pilot placements: • LOC (fully local): T2 on R1 , T3 on R2 ; • SO (static offload): T2 and T3 on E; • HYB (hybrid): T2 on E, T3 on R2 . The window-level placement problem is πk∗ ∈ arg

min Jk (π; πk−1 ),

π∈Πcand

(7)

subject to L95,k (π) ≤ Lmax and per-node utilisation caps Un,k (π) ≤ Umax for all n ∈ N .

IV. DYNAMIC TASK P LACEMENT A LGORITHM A. Window-Based DTP Procedure Algorithm 1 describes the DTP procedure running on the edge node – E. Two parameters jointly enforce hysteresis: • Nmin : Minimum dwell windows; prevents rapid flapping when metrics fluctuate near thresholds; • ∆min : Cost-improvement threshold; ensures only statistically significant improvements trigger migration. Denote that per-window complexity is O(|Πcand |), which is always lightweight on any edge node if task placements occur.

Algorithm 1 Window-Based Dynamic Task Placement (DTP) Require: W ; (αL , αV , αr , αe , αS ); ∆min ; Nmin ; Πcand 1: Init π0 ∈ Πcand ; dwell nd ← 0 2: for k = 1, 2, . . . do 3: Apply πk−1 for W cycles; collect {L95,k , VD,k , Ur,k , Ue,k } 4: Compute Jk (πk−1 ; πk−2 ) via (6) 5: if nd < Nmin then 6: nd ← nd +1; πk ← πk−1 ; continue {enforce min. dwell} 7: end if 8: Estimate Jk (π; πk−1 ) for all π ∈ Πcand 9: πk∗ ← arg minπ Jk (π; πk−1 ) 10: ∆J ← Jk (πk−1 ; πk−2 ) − Jk (πk∗ ; πk−1 ) 11: if ∆J > ∆min then 12: πk ← πk∗ ; nd ← 0 {migrate} 13: else 14: πk ← πk−1 ; nd ← nd +1 {hold} 15: end if 16: end for B. Estimating QoS for Non-Active Placements Scoring placements not currently active requires estimating their idle QoS metrics. The three initial complementary mechanisms are supported along the operation loop as follows: • Static profiling: Compute times cv (n) and communication delays measured offline and stored as lookup tables. • Online emulation: A small fraction of cycles is routed through each alternative in the background to update estimates continuously during the non-placement stage. • Conservative scaling: Upper-bound estimates derived from current measurements or passive monitored feedback and known compute/network ratios. The original algorithm is agnostic to the choice, provided cost comparisons remain qualitatively consistent for DTP. V. Q UALITATIVE A NALYSIS AND D EPLOYMENT C ONTEXT A. Anticipated Placement Trajectories Table I summarizes the expected steady-state DTP behaviors relative to LOC and SO baselines across four pilot canonical stress scenarios. The policy converges toward LOC under adverse network conditions and toward SO under robot CPU stress; HYB dominates when compute and communication resources are jointly constrained, by offloading the heavier T2 (Perception) to E while keeping T3 (Planning) on R2 . B. Co-Design Implications and the PAI Nervous System Cast as the CFA 3C co-design principle-based orchestration layer, the DAG-based QoS-aware DTP framework provides three concrete guarantees against the failure modes intrinsic to networked factory control following industrial standards: • The hard-anchoring of T1 and T4 preserves the submillisecond reflex loop on local hardware, decoupling actuation safety and dependencies from network state;

The relocatable adaptation space accommodates the heterogeneous compute demands of T2 and T3 , including emerging VLA/LLM workloads, without overloading machine hardware or hard-coding offload decisions; • The Hamming-distance switching penalty in (5) provides a mathematically bounded guarantee against placement chatter, translating directly to operator-observable control-loop stability and to the deterministic timing budgets expected of TSN-class substrates [5]. Technically, a factory automation system that cannot guarantee VD ≤ 5% across the deployment envelope cannot credibly claim Industry 4.0 production maturity [9] toward industrial agentic automation for Industry 5.0 readiness. •

VI. E MPIRICAL S TRATEGY AND VALIDATION ROADMAP Validation proceeds in two main phases, each with concrete parameters drawn directly from the physical robotic testbed to ensure direct comparability and reflection on the previous empirical LOC, SO, and ATP baselines. A. Phase 1 – Discrete-Event Simulation A discrete-event simulator parameterized will validate costfunction sensitivity and hysteresis bounds against the predictions in Table I. Key parameters are: • Control period P ∈ [20, 50] ms with deadline D ≤ P ; • Baseline link RTT of 1–2 ms representative of an engineered industrial LAN, with TSN-class jitter envelopes as a sensitivity axis following standardization [5]; • CPU stress profiles applied to the R1 -equivalent node; • Gaussian network fault injection (e.g., µ = 25 ms, σ = 5 ms, ploss = 2%), plus additional extended verification. B. Phase 2 – Hardware-in-the-Loop (HIL) The modular cloud-edge-robotics testbed will provide: • Measured compute profiles cv (n) on each physical node; • Empirical network delay distributions on the R1 –E and E–R2 links with extended IIoT devices/PLCs; • Empirical CDFs of L95 , VD , Ur , and Ue for LOC, SO, ATP, and DAG-QoS DTP across all stress scenarios. Phase 2 enables direct quantitative comparison between the proposed multi-stage DTP and the single-stage ATP baseline. VII. C ONCLUSIONS AND F UTURE W ORK In this WiP paper, we presented the DAG-based QoSaware DTP, the formal multi-objective extension of our edgebased QoS-aware ATP framework. By formalizing the visualservoing pipeline as the DAG, introducing the window-based cost function with the Hamming-distance switching penalty, and generalizing the ATP algorithm to concurrent multi-task placement, the extended framework addresses three structural limitations of single-task adaptive placement: limited scope, threshold-based decisions, and the absence of a formal switching cost model. The result is a runtime determinism layer compatible with the CFA-based 3C co-design ethos central to factory communications and mature industrial automation.

TABLE I E XPECTED DOMINANT DTP PLACEMENT PER STRESS SCENARIO . Scenario

LOC

SO

Baseline

Low VD ; high Ur

Robot CPU stress Edge CPU stress Network impairment

High VD (Ur sat.) Stable; Ur elevated Low variance

Low Ur ; moderate VD Mitigated; higher L95 High VD (Ue sat.) High VD via jitter

DTP (expected) Mostly LOC; occasional HYB Converges to SO LOC or HYB Converges to LOC

Beyond factory robotic automation, this middle layer will act as one of the key cores in termed the “PAIs Nervous System”: the runtime infrastructure that lets LLM agents and VLA models operate reliably under real compute and network variability on different PAIs and nearby edges/clouds. Future work will extend the proposed framework in this WiP towards full deployment through: (i) empirical validation on the physical testbed; (ii) expansion of the candidate placement set beyond LOC/SO/HYB and extension to multi-robot fleets where a single edge serves several DTP loops concurrently; (iii) online cost-weight adaptation of hyperparameters via Bayesian optimization or lightweight meta-learning; (iv) joint compute–network co-scheduling with deterministic substrates, elevating the network from a delay distribution to a cooptimized hardware resource, and considering new QoS dimensions such as Age-of-Information (AoI). R EFERENCES [1] R. Li, Y. Zhou, Y. Zhu, K. Chen, J. Wang, S. Wang, K. Hu, M. Yu, B. Jiang, Z. Su, J. Ma, X. He, Y. Shen, Y. Yang, G. Ren, M. Yao, W. Wang, and Y. Mu, “RoboClaw: An agentic framework for scalable long-horizon robotic tasks,” 2026. [2] H. Li, X. Cao, Z. Zeng, Y. Wu, Y. Zhang, and Y. Xia, “OpenGo: An OpenClaw-based robotic dog with real-time skill switching,” 2026. [3] J. Jin, K. Yu, J. Kua, N. Zhang, Z. Pang, and Q.-L. Han, “Cloud-fog automation: Vision, enabling technologies, and future research Directions,” IEEE Trans. Ind. Inform., pp. 1–16, 2023. [4] J. Jin, Z. Pang, J. Kua, Q. Zhu, K. H. Johansson, N. Marchenko, and D. Cavalcanti, “Cloud-fog automation: The new paradigm towards autonomous industrial cyber-physical systems,” IEEE J. Sel. Areas Commun., 2025. [5] T. Zhang, G. Wang, C. Xue, J. Wang, M. Nixon, and S. Han, “TimeSensitive Networking (TSN) for industrial automation: Current advances and Future directions,” ACM Comput. Surv., vol. 57, no. 2, Oct. 2024. [6] T. Luu, Q. Nguyen, T. Tran, M. Tran, S. Ding, J. Kua, and T. Hoang, “Enhancing real-time robot teleoperation with immersive virtual reality in industrial IoT networks,” Int. J. Adv. Manuf. Technol., vol. 139, no. 11, pp. 6233–6257, Aug. 2025. [7] Z. Tong, J. Deng, J. Mei, Y. Zhang, and K. Li, “Multi-objective DAG task offloading in MEC environment based on federated DQN with automated hyperparameter optimization,” IEEE Trans. Serv. Comput., vol. 17, no. 6, pp. 3999–4012, 2024. [8] J. Liu, J. Ren, Y. Zhang, X. Peng, Y. Zhang, and Y. Yang, “Efficient dependent task offloading for multiple applications in MEC-cloud system,” IEEE Trans. Mobile Comput., vol. 22, no. 4, pp. 2147–2162, 2023. [9] R. Andreoli, R. Mini, P. Skarin, H. Gustafsson, J. Harmatos, L. Abeni, and T. Cucinotta, “A multi-domain survey on time-criticality in cloud computing,” IEEE Trans. Serv. Comput., vol. 18, no. 2, pp. 1152–1170, 2025.