Paper deep dive
Modeling and Validation of Quality of Control for Edge-Offloaded Collaborative Navigation
Neelabhro Roy, Mikael Hammarling, Victor Nan Fernandez-Ayala, Gourav Prateek Sharma, Mani H. Dhullipalla, Dimos V. Dimarogonas, James Gross
Intelligence
Status: succeeded | Model: Gemma-4-26B-A4B | Prompt: intel-v1 | Confidence: 90%
Last extracted: 7/17/2026, 2:16:09 AM
Summary
This paper models and validates a Quality of Control (QoC) framework for edge-offloaded collaborative navigation of non-holonomic AGVs over a private 5G network. It quantifies how stochastic wireless delay and reliability variations impact closed-loop performance and energy efficiency, demonstrating through simulation and a 5G testbed that ROS 2 RELIABLE QoS outperforms BEST-EFFORT by 51.5% in QoC.
Entities (8)
Relation Signals (6)
Private 5G Testbed → introduces → Stochastic Wireless Delay
confidence 95% · shift from rigid, wired industrial setups to adaptive wireless solutions introduces significant stochasticity in the form of delay, jitter, and packet loss
ROS 2 QoS Policies → outperforms → BEST-EFFORT QoS
confidence 95% · RELIABLE QoS offers 51.5% better QoC than BEST-EFFORT under certain experimental settings
Quality of Control (QoC) → models → Closed-loop Performance
confidence 90% · Our approach (i) models end-to-end network effects on closed-loop performance
TurtleBot 3 → operateson → Private 5G Testbed
confidence 90% · Our experiments use N=2 TurtleBots... operating on a private 5G testbed in the KTH R1 hall
Edge Computing → offloads → Control Tasks
confidence 85% · offloading diverse workloads, ranging from perception to motion planning for industrial automated guided vehicles (AGVs) and collaborative robots to edge/cloud systems
Nav2 → utilizes → MPPI and A* Planner
confidence 85% · this setup utilizes a hierarchical architecture: an A* global planner [19] for path generation and a rolling-horizon MPPI local controller for tracking
Cypher Suggestions (0)
No Cypher suggestions yet.
Abstract
Abstract:Collaborative control in complex environments is severely challenged by stochastic wireless delay and reliability variations, which can degrade navigation, tracking, and collision avoidance. These network-induced uncertainties complicate the maintenance of energy efficiency during collaborative tasks, and can potentially lead to over-provisioning of resources. In this paper, for a navigation setup with dynamic collision avoidance, we address this challenge by expanding the quality of control (QoC) framework from prior works to practical robotic models. Our approach (i) models end-to-end network effects on closed-loop performance, (ii) systematically explores the impact of various control parameters dictating robotic motion on network latency-reliability (iii) validates these models through experiments on a private 5G testbed across varying delay, reliability and control configurations. Our analysis indicates the optimal control-communication co-design operating regimes for practical robots and also compares the QoC performance of standard ROS~2 quality of service (QoS) policies under real-world conditions and showing how RELIABLE QoS offers 51.5% better QoC than BEST-EFFORT under certain experimental settings.
Tags
Links
- Source: https://arxiv.org/abs/2607.14853v1
- Canonical: https://arxiv.org/abs/2607.14853v1
Trouble viewing inline? Open PDF directly →
Full Text
40,648 characters extracted from source content.
Expand or collapse full text
Modeling and Validation of Quality of Control for Edge-Offloaded Collaborative Navigation Neelabhro Roy2, Mikael Hammarling2, Victor Nan Fernandez-Ayala2, Gourav Prateek Sharma1, Mani H. Dhullipalla2, Dimos V. Dimarogonas2 and James Gross2 Abstract Collaborative control in complex environments is severely challenged by stochastic wireless delay and reliability variations, which can degrade navigation, tracking, and collision avoidance. These network-induced uncertainties complicate the maintenance of energy efficiency during collaborative tasks, and can potentially lead to over-provisioning of resources. In this paper, for a navigation setup with dynamic collision avoidance, we address this challenge by expanding the quality of control (QoC) framework from prior works to practical robotic models. Our approach (i) models end-to-end network effects on closed-loop performance, (i) systematically explores the impact of various control parameters dictating robotic motion on network latency-reliability (i) validates these models through experiments on a private 5G testbed across varying delay, reliability and control configurations. Our analysis indicates the optimal control-communication co-design operating regimes for practical robots and also compares the QoC performance of standard ROS 2 quality of service (QoS) policies under real-world conditions and showing how RELIABLE QoS offers 51.5% better QoC than BEST-EFFORT under certain experimental settings. I Introduction With the emergence of cloud and edge computing promising low latency and ultra reliable wireless communications, several works have investigated offloading diverse workloads, ranging from perception to motion planning for industrial automated guided vehicles (AGVs) and collaborative robots to edge/cloud systems [1, 11, 12]. Such architectures enable advanced capabilities like external sensing, centralized mapping, and intensive compute offloading, thereby reducing the onboard compute burden while broadening robotic capabilities, important for flexible automation and production [12, 17, 4]. However, the shift from rigid, wired industrial setups to adaptive wireless solutions introduces significant stochasticity in the form of delay, jitter, and packet loss [12, 10]. These uncertainties directly affect sensing, navigation, and, most critically, the stability of control loops in coordination or collaboration [12, 13]. Effective collaborative task execution remains a challenging problem and often relies on a centralized task planner in an edge server responsible for mapping, path planning, and navigation [6]. A fundamental limitation and design parameter in these deployments is the robotic energy budget, due to battery constraints, which further warrants efficient path planning to prevent energy waste. Moreover, managing energy expenditure is not merely a hardware concern originating from battery issues, but a crucial constraint for wireless collaborative robotics (WCRs) [12] in general. As control offloading to the edge increases, managing communication-induced uncertainty becomes vital for state and energy regulation [12]. While energy-centric studies [3, 2] optimize search-and-rescue efficiency assuming the control loop is less time-sensitive and robust to typical 5G latencies, they overlook how stochastic delay and reliability shape closed-loop coordination. Recent distributed evaluations, such as the robot operating system (ROS 2)-based 5G image-processing stack in [17], highlight the complex factors influencing reaction times but rely on simple safety timeouts to mitigate jitter. Existing literature [17, 8, 14] largely analyzes isolated parameters like sampling rate or latency for individual robots. Consequently, there is still a lack of a comprehensive abstraction that jointly accounts for communication and control parameters to quantify collaborative system performance and energy expenditure. Quality-of-Control (QoC)-based metrics and abstractions have been studied in recent works [12, 10, 13] linking wireless parameters to robotic behaviour. By unifying energy and control expenditure into a single objective, QoC further enables energy gains by optimizing the delay-reliability trade-off as shown in [13]. However, this approach was limited to simulations only, considered simplified consensus dynamics and lacked experimental validation in [12, 13]. Crucially, it overlooked non-holonomic AGV dynamics, where the non-linear coupling of orientation and translation makes tracking more sensitive to network stochasticity than simpler linear, holonomic models [16]. To address these gaps, this work develops a system model tailored for practical robotic motion and validates its effectiveness through both extensive simulations and practical 5G testbed experiments. Our main contributions are summarized as follows: • Non-Holonomic System Modeling: We generalize the frameworks in [12, 13] by incorporating non-holonomic kinematic constraints. Unlike previously considred holonomic systems, which can translate instantaneously in any direction, our proposed model accounts for motion restricted by the robot’s current heading. This enables the characterization of network stochasticity on coupled translational and rotational AGV dynamics, together. • QoC-Based Analysis: We propose and derive a QoC abstraction for 5G-Edge offloaded navigation featuring dynamic collision avoidance, quantifying the specific impact of various control parameters and network-induced impairments on system performance, towards co-design. • Experimental Validation: We validate our simulation framework using a private 5G testbed [7] and the Nav2 architecture. The experimental results confirm the framework’s utility, further demonstrating that ROS 2 RELIABLE QoS yields a 51.5% higher QoC compared to BEST-EFFORT under real-world settings. I System Model In this work, we model a WCR system, illustrated in Fig. 1, which comprises robots collaborating over a shared 5G wireless network. This section discusses key system aspects, including the robot dynamics and navigation, aspects of the 5G network forming the underlying communication channel, and finally, the problem statement. I-A Overview and Probem Statement Offloading control to the 5G edge enables advanced collaboration but introduces stochastic uncertainties that can compromise stability and system performance. Existing frameworks often overlook the non-linear coupling of orientation and translation in practical AGVs, leaving a gap between theoretical models and real-world navigation. This paper addresses this by modeling and validating a QoC-based abstraction for non-holonomic motion to identify optimal control-communication regimes for co-design. We consider N AGVs (TurtleBot 3111https://w.turtlebot.com/) performing navigation with dynamic collision avoidance via an edge-based controller. Because the robots collaborate exclusively through this centralized offloading rather than peer-to-peer links, our framework explicitly models the stochastic delay and reliability of the periodic, end-to-end information loop over a private 5G network [7]. Figure 1: Overview of a WCR setup with control offloaded to a centralized server running Nav2. TUL,iT_UL,i, TC,iT_C,i and TDL,iT_DL,i have been defined in Section I-D. Figure 2: Illustration of the costmap utilized by two different robots for Nav2-based navigation. I-B Robot Dynamics and Control I-B1 Modeling robot dynamics We extend the simulation framework discussed in [12, 13] towards modeling of AGV motion observed in practice. We consider planar and angular motion of robots i∈1,…,Ni∈\1,…,N\ under an AGV kinematic model [16] with state and commanded inputs: i(t)≜[xi(t)yi(t)θi(t)]⊤∈ℝ3,x_i(t) [\,x_i(t)\;\;y_i(t)\;\; _i(t)\,] ^3, (1) i(t)≜[νi(t)ωi(t)]⊤∈ℝ2,u_i(t) [\, _i(t)\;\; _i(t)\,] ^2, (2) where (xi,yi)(x_i,y_i) is the planar position, θi _i is the heading angle, and (νi,ωi)( _i, _i) denote the translational and the rotational speeds, respectively. The state evolution [16] is given by: ˙i(t)=J(θi(t))i(t),J(θ)≜[cosθ0sinθ001], x_i(t)=J( _i(t))\,u_i(t), J(θ) bmatrix θ&0\\ θ&0\\ 0&1 bmatrix, (3) i.e., x˙i=νicosθi x_i= _i _i, y˙i=νisinθi y_i= _i _i, and θ˙i=ωi θ_i= _i. Let i(t)≜[xi(t)yi(t)]⊤p_i(t) [\,x_i(t)\;\;y_i(t)\,] . At discrete sampling instants tp=t0+pτt_p=t_0+pτ, where t0t_0 is the initial time, p is the sampling index, and τ is the fixed sampling period, the edge-based controller (Fig. 1) computes the control command, applied via a zero-order hold (ZOH), such that i(t)x_i(t) evolves as: ˙i(t)=J(θi(t)),i(tp),t∈[tp,tp+1), x_i(t)=J( _i(t)),u_i(t_p), t∈[t_p,t_p+1), (4) where J(θi)J( _i) is the Jacobian mapping commanded velocities to the global frame, as shown in Eq.(3). I-B2 Control commands In this paper, the objective of the robots in the system is to arrive at their pre-specified destinations by successfully avoiding collisions. To facilitate this, prior to designing appropriate control commands ui(tp)u_i(t_p) in Eq.(4), we define the following variables: At sampling instants tpt_p, the edge controller computes the desired planar velocity ides(tp)v_i^des(t_p) for each robot i moving towards the target coordinate ⋆∈ℝ2p ^2 using estimated positions ^i(tp) p_i(t_p): ides(tp)≜−κ(^i(tp)−⋆)+∑j≠ijrep(tp),v_i^des(t_p) -κ ( p_i(t_p)-p )+ _j≠ iv^rep_ij(t_p), (5) where κ>0κ>0 is the goal gain capturing the strength of the closed-loop response, also absorbing unmodeled nonlinearities222Velocity commands are generated by the Nav2 stack (A*/MPPI) running on the edge. In simulation, we can use the analytical feedback law in this section as a surrogate controller to isolate QoC sensitivity to delay/reliability while matching the measured timing and loop-completion statistics., and ijrepv^rep_ij is a repulsive collision-avoidance term, tackling potential collisions between robots i and j. Note that Eq.(5) is inspired from robot control for obstacle avoidance using artificial potential fields (APFs) [15]. Let dij(tp)≜‖^i(tp)−^j(tp)‖2d_ij(t_p) \| p_i(t_p)- p_j(t_p)\|_2 be the estimated inter-robot distance and let dsafeeffd_safe^eff be the effective safety radius that expands with information staleness: dsafeeff(tp)≜dsafe+cmissmax(0,max(ai(tp),aj(tp))−Δ),d_safe^eff(t_p) d_safe+c_miss (0, (a_i(t_p),a_j(t_p))- ), (6) where dsafe>0d_safe>0 is the nominal safety distance, cmiss≥0c_miss≥ 0 is a penalty coefficient, and Δ is the loop completion deadline. The age variable ai(t)a_i(t) captures the time since the last successful control update (Eq.(16)). This keeps robots from violating a minimum distance [15], and becomes more conservative under packet drops via Eq.(6). For a repulsion gain η>0η>0: ijrep(tp)≜η(1dij−1dsafeeff)^i(tp)−^j(tp)dij3,dij<dsafeeff,0,otherwise.v^rep_ij(t_p) casesη ( 1d_ij- 1d_safe^eff ) p_i(t_p)- p_j(t_p)d_ij^3,&d_ij<d_safe^eff,\\[4.0pt] 0,&otherwise. cases (7) To accommodate AGV kinematics, idesv_i^des is mapped to translational and rotational speeds (vi,ωi)(v_i, _i). The desired heading angle θides(tp) _i^des(t_p) and the orientation error eθ,i(tp)e_θ,i(t_p) relative to the estimated heading θ^i(tp) θ_i(t_p) are given by: θides(tp)≜atan2(vy,ides(tp),vx,ides(tp)) _i^des(t_p) 2\! (v^des_y,i(t_p),\,v^des_x,i(t_p) ) (8) eθ,i(tp)≜wrap(θides(tp)−θ^i(tp))∈[−π,π].e_θ,i(t_p) \! ( _i^des(t_p)- θ_i(t_p) )∈[-π,π]. (9) Eq.(8) computes a desired heading angle that points in the direction of the desired planar velocity, while Eq.(9) compares the desired heading to the currently known heading estimate, and the wrapwrap function ensures the shortest rotation direction. A proportional heading controller with gain γ>0γ>0 converts the desired velocity into translational and rotational commands: νi(tp) _i(t_p) ≜‖ides(tp)‖2max(cos(eθ,i(tp)),0), \|v_i^des(t_p)\|_2 ( (e_θ,i(t_p)),0), (10) ωi(tp) _i(t_p) ≜γeθ,i(tp). γ\,e_θ,i(t_p). The applied input i(tp)=sat∞(~i(tp),max)u_i(t_p)=sat_∞( u_i(t_p),u ) is finally subject to componentwise velocity saturation with caps max≜[vmax,ωmax]⊤u [v ,ω ] (like [13]) and saturation law: sat∞(,max)≜[max(0,min(ν,νmax))max(−ωmax,min(ω,ωmax))]sat_∞(u,u ) bmatrix (0,\, (ν,\,ν ))\\ (-ω ,\, (ω,\,ω )) bmatrix (11) I-C Navigation Building on the simulation system model in Section I-B, we adopt the Nav2 and model predictive path integral (MPPI) stack [18] as a more general representative candidate for real experimental validation. While the preceding framework can generalize to various planning and control algorithms, this setup utilizes a hierarchical architecture: an A* global planner [19] for path generation and a rolling-horizon MPPI local controller for tracking. Both components rely on a multi-layer costmap [5] (Fig. 2) with a static layer shown in black for known, time-invariant structures of the environment (e.g., walls and permanent obstacles) derived from a prebuilt map; a light blue dynamic layer for obstacles detected online with a LiDAR; and lastly, an inflation layer in pink that expands obstacle boundaries for extra safety using the robot footprint, localization uncertainty, and network-induced control delays. Formally, the workspace is discretized into a 2D grid, with j denoting the direction in the x-dimension and k in the y-dimension. Each cell jkjk stores an aggregated cost cjk∈ℝ≥0c_jk _≥ 0: cjk=cjkstat+cjkobs+cjkinfl,c_jk=c^stat_jk+c^obs_jk+c^infl_jk, (12) with the superscripts denoting the static, obstacle, and inflation layer costs, respectively. The global planner treats this grid as a graph, computing paths that minimize cumulative cost. Similarly, MPPI evaluates candidate rollouts over horizon ℋH by accumulating state-dependent costs: Jmap=∑t∈ℋc((t)).J_map= _t c (x(t) ). (13) This unified representation facilitates the integration of spatial risk and safety margins across both planning and control layers. In contrast to conventional Nav2 architectures, which typically employ distinct local and global costmaps, we adopt a single, shared map representation. While the current study does not exhaustively analyze these costmaps, this consolidated structure provides a formal framework for future research into network-aware navigation. I-D Communication model Each control/navigation update described in Section I-B and I-C traverses a 5G network with a stochastic loop time Tnet,iT_net,i, comprising uplink (TUL,iT_UL,i), compute (TC,iT_C,i), and downlink (TDL,iT_DL,i) stages, as described here: Tnet,i≜TUL,i+TC,i+TDL,i.T_net,i T_UL,i+T_C,i+T_DL,i. (14) For a deadline Δ , we define the loop completion probability: pinet(Δ)≜ℙ(Tnet,i≤Δ)=Fnet,i(Δ),p^net_i( ) (T_net,i≤ )=F_net,i( ), (15) where Fnet,iF_net,i is obtained from measured/simulated compute distributions (described as compute-time model in Table I). If the loop completes in time, the updated control applies to both axes; otherwise, the input stays equal to the previous sample: i(tp)=inew(tp),Tnet,i≤Δi(tp−1),otherwise.u_i(t_p)= casesu^new_i(t_p),&T_net,i≤ \\ u_i(t_p-1),&otherwise. cases (16) This implicitly affects QoC computation discussed in Section I-E. I-E Quality of Control To quantify the QoC for the navigation task, we define a nominal, time-varying reference trajectory ¯(t) p(t). This reference trajectory represents the unconstrained optimal path to the destination, computed via Eq.(5) by neglecting obstacles. The control performance is then characterized by the tracking error or disagreement i(t) δ_i(t), representing the state deviation from this nominal trajectory: i(t)≜i(t)−¯(t). δ_i(t) _i(t)- p(t). (17) We define QoC from this disagreement energy. Let W=diag(wx,wy)≻0W=diag(w_x,w_y) 0 be an optional axis-weighting matrix. Define the cost for robot i over horizon [0,T][0,T]: JiC≜∫0T‖i(t)‖W2t,‖i‖W2=i⊤Wi.J^C_i _0^T\!\| δ_i(t)\|_W^2\,dt, \| δ_i\|_W^2= δ_i W δ_i. (18) The normalized cost or the area under the disagreement curve (AUC), denoted by Jnorm,iCJ^C_norm,i, is defined as: Jnorm,iC≜∫0T‖i(t)‖W2‖i(0)‖W2t.J^C_norm,i _0^T\! \| δ_i(t)\|_W^2\| δ_i(0)\|_W^2\,dt. (19) We map this to a bounded QoC, defined by QiQ_i, using: Qi≜11+min(Jnorm,maxC,Jnorm,iC)JmedC,Qi∈[0,1],Q_i 11+ (J^C_norm, ,\;J^C_norm,i)J^C_med, Q_i∈[0,1], (20) where Jnorm,maxCJ^C_norm, is a scenario-dependent reference maximum ensuring convergence across experiments and JmedCJ^C_med is the median AUC value. Since i(t) δ_i(t) depends on the applied saturated control and on missed updates through ZOH, QiQ_i captures the effects of saturation maxu , delay, and reliability. Please note that this QoC measure is an energy-relevant proxy rather than a direct battery measurement. Larger disagreement, detours, oscillatory corrections, and stale-control recovery typically increase motion and control effort, however, direct validation against battery usage is left for future work. I QoC Abstraction for AGV motion We begin this section by detailing how a QoC model could be obtained, given a system of N AGVs. The objective is to observe how Eq.(18) can help us quantify system performance with respect to various control and communication parameters together. To realize this, we simulate the offloaded control loop where each robot evolves according to the AGV model ˙i(t)=J(θi(t))i(t) x_i(t)=J( _i(t))\,u_i(t) with i(t)=[νi(t)ωi(t)]⊤u_i(t)=[ _i(t)\; _i(t)] and ZOH updates at tp=t0+pτt_p=t_0+pτ, with the parameters described in Table I. Please note that N=2N=2 was simulated to match the experimental setup and the abstraction can readily scale for a higher number of robots [13]. For larger fleets, the aggregate network demand scales with the number of periodic state and control command streams, while the edge compute component may increase due to multi-robot planning, costmap updates, collision checking and queueing effects. Thus, scaling to multiple AGVs would primarily alter the empirical distribution of Tnet,iT_net,i in Eq.(15) and the required network resource provisioning, which is precisely the type of shift captured by the QoC abstraction through pinet(Δ)p_i^net( ). Fig. 3 shows how AUC (Jnorm,iCJ^C_norm,i) varies versus delay-reliability together for selected parameters η and τ. (a) Repulsive Gain. (b) Sampling period. Figure 3: AUC variation for selected control parameters Here, we can observe that at very low latency-reliability levels, AUC and thus the system performance and energy expenditure worsens substantially, across the entire range of simulated values for both the repulsive gain and the sampling period. This is attributed to very high packet drops that did not let the control commands be updated. Once sufficient reliability is achieved, AUC remains largely unchanged with different sampling periods. On the contrary, we still see worsening AUC as the repulsive gain grows together with high delays. TABLE I: Simulation parameters Parameter Value Number of robots N=2N=2 Simulation horizon Tsim=30sT_sim=30~s Sampling period (ZOH) τ=10msτ=10~ms Compute-time model TruncNormal(μc,σc)( _c, _c) on [0,∞)[0,∞) Compute-time mean/std μc=0.20s _c=0.20s, σc=0.05s _c=0.05s Goal and Repulsion gain κ=2κ=2, η=0.6η=0.6 Speed caps νmax=0.7ν =0.7, ωmax=1.4ω =1.4 IV Experimental Validation In this section, we describe how we perform experiments with the system model in Section I and then perform a limited validation of the abstraction in Section I. IV-A Testbed and Experimental parameters Our experiments use N=2N=2 TurtleBots (due to spatial, hardware and ROS 2 data distribution service (DDS) scaling constraints [9]) operating on a private 5G testbed in the KTH R1 hall, with edge computation [7]. The navigation stack for these robots runs on the edge system, and all communication occurs over the 5G network, operating on band n78. The TurtleBots were connected to the private 5G network through an external 5G dongle attached to the robot-side compute/network interface as shown in Fig. 4. Figure 4: Private 5G testbed with TurtleBots and 5G dongles. Fig. 5 denotes how the robots are tasked with a navigation objective with dynamic collision avoidance towards destination coordinate p⋆∈ℝ2p ^2. For each run, we log: (i) commanded velocities, (i) ROS 2 QoS settings, and (i) timestamps at UL send, edge receive, edge send, and DL receive for all control-relevant messages. To evaluate performance, i(t)p_i(t) takes the pose estimate for robot i∈1,…,Ni∈\1,…,N\, from the robot’s adaptive Monte Carlo localisation (AMCL) pose topic. We treat the AMCL pose as the best estimate of robot position in the global frame for computing disagreement-based metrics. IV-B Validation methodology To emulate various delay-reliability combinations towards validating the abstraction simulations in Section I, we inject an artificial delay until DcfgD_cfg(ms) into the velocity command stream, generated by the edge. This utilizes robot-local timestamping, eliminating the need for clock synchronization between the robot and the edge server: • At time tst_s on the robot clock, a source state message generated on the robot is transmitted into the navigation pipeline (where robot i′si s AMCL pose is computed). • The edge compute-based navigation stack produces a velocity command corresponding to that state. • A delay node buffers the command totalling a configured duration DcfgD_cfg and only then publishes the delayed command that is actually applied by the robot base (Eq.(16)). IV-C Experimental Validation In this subsection we discuss the assumptions made and the simulation settings based on the description in Section I to match the subsequent experimental settings. Let ts,i(k)t_s,i^(k) be the generation time of the kthk^th source state sample for robot i (robot clock), i.e., the timestamp embedded in the source message header: ts,i(k)≜stamp(i(t)(k)).t_s,i^(k)\; \; stamp (x_i(t)^(k) ). Let ta,i(k)t_a,i^(k) be the time at which the corresponding delayed command is applied on the robot: ta,i(k)≜record(i(k))t_a,i^(k)\; \; record (u_i^(k) ), and we recover ts,i(k)t_s,i^(k) from the propagated header stamp of the delayed command. The configured delay is then defined as: Dcfg(k)≜ta,i(k)−ts,i(k).D_cfg^(k)\; \;t_a,i^(k)-t_s,i^(k). Provided the configured delays, and the simulation settings from Section I, let S(φ)S( ) denote the simulated AUC at configured delay φ∈ and let E(φ)E( ) denote the experimentally measured AUC at the same φ . For comparison between these two metrics, we first normalize them using Eq.(21) and repeat the equivalent for E(φ)E( ): S(φ)= 100⋅S(φ)−minφ′∈S(φ′)maxφ′∈S(φ′)−minφ′∈S(φ′),S( )\;=\;100· S( )- _ S( ) _ S( )- _ S( ), (21) Figure 5: Adopted trajectory for a representative run of two Turtlebots performing navigation with 5G-Edge Controllers. Figure 6: Normalized simulation-based and experimental observations on the AUC variation versus delays. Fig. 6 plots how the experimental and the simulated AUC values (as in Eq.(21)) vary across different delay-reliability pairs. Both cases follow similar trends, consistent with findings in [12]. This alignment is reinforced further by how the reliability-vs-delay transition happens in both simulation and experiment at similar delays (e.g., reliability improving strongly around 100–200 ms), and the delay dominance region starts at similar delays as well. The observed mismatch between simulation and experiments should be interpreted in light of the different roles of the collision-avoidance mechanism in the two cases. In simulation, the repulsion gain ηrep _rep in Eq.(7) provides a compact surrogate for inter-robot collision avoidance, allowing us to isolate how delay, reliability, and control parameters affect QoC. In the experimental Nav2/MPPI stack, however, collision avoidance is not governed by a single scalar gain, and instead emerges from the interaction of obstacle-layer updates, inflation costs, MPPI rollout costs, and local path-selection effects. Hence, varying ηrep _rep experimentally would require retuning several Nav2/MPPI and costmap parameters rather than changing a single physical parameter. The comparison in Fig. 6 is therefore intended as a trend-level validation of the QoC abstraction, showing that the low-reliability, intermediate-reliability, and delay-dominant regimes are reproduced, while exact AUC matching is limited by the richer environment-dependent behavior of the real navigation stack. V Evaluations In this section, we first describe how the QoC abstraction simulations reveal the behavior of different parameters for AGV motion and indicate the optimal operating regimes across different parameters and then anlayze QoC experimenatlly for ROS-based reliability QoS settings. Figure 7: Parameter sweep across delay-reliability regions V-A Impact of various parameters on QoC-Simulations In this subsection, we explore how varying different parameters crucial for AGV motion across various delay-reliability regions impacts QoC, via simulations. We further indicate optimal operating regimes towards co-design aimed at maximizing system performance and minimizing energy and state expenditure. For each design parameter Θ∈νmax,ηrep,τ,κ,γ,cmiss ∈\ν , _rep,τ,κ,γ,c_miss\ we perform a one-at-a-time sweep over a fixed grid ΛΘ _ while keeping all other parameters at their baseline values. For every pair (Θ,Δ)( , ) on the delay axis Δ∈[0,800] ∈[0,800]\,ms, assuming enough network resources or 5G physical resource block (PRBs) are available, we evaluate JnormCJ^C_norm and compute QoC. We compare three policies per sweep: (i) Co-design selects optimal pair (Θ⋆,Δ⋆)=argmaxΘ∈ΛΘ,ΔQ(Θ,Δ)( , )= _ ∈ _ , Q( , ); (i) MinDelay fixes Δ=Δmin = _ and tunes Θ only for that regime, ΘD⋆=argmaxΘ∈ΛΘQ(Θ,Δmin) _D = _ ∈ _ Q( , _ ); and (i) MaxRel fixes Δ=Δmax = _ (highest loop completion probability) and tunes Θ only for that regime, ΘR⋆=argmaxΘ∈ΛΘQ(Θ,Δmax) _R = _ ∈ _ Q( , _ ). The grouped bar chart in Fig. 7 reports the achieved QoC at the chosen operating point per sweep. It shows MinDelay can be suboptimal because low delay does not guarantee timely/reliable loop completion, leading to stale/missed updates (ZOH) and increased disagreement; MaxRel improves delivery probability but incurs large delays, making the closed-loop response sluggish. In contrast, the co-design policy selects intermediate delay-reliability regimes and corresponding parameter values that maximize QoC, consistently, highlighting the need for control–communication co-design rather than optimizing delay or reliability in isolation. V-A1 Parameter-specific observations (i) νmaxν governs how aggressively robots can react to goal and avoidance commands; under MinDelay (low reliability), higher νmaxν amplifies the effect of stale headings and can increase disagreement, whereas co-design prefers moderate delays where updates are reliable enough to safely exploit higher mobility. (i) ηrep _rep trades cohesion against safety: larger repulsion spreads robots apart and can increase disagreement, so the QoC-optimal setting tends to be moderate (or lower), while still preventing close approaches. (i) τ changes both the control update rate and the per-attempt transmission granularity; when reliability is low, frequent updates do not help because many are missed (ZOH dominates), but at reliable operating points an appropriate τ reduces staleness without inducing overly sluggish behavior. (iv) κ sets the strength of goal attraction; too large κ under stale information can cause overshoot/oscillation (worse QoC), while too small κ yields slow convergence, so co-design selects κ jointly with an intermediate delay where the loop is both responsive and reliable. (v) γ controls heading correction; aggressive heading gains are sensitive to delayed direction estimates, so co-design improves QoC by operating in regimes where heading errors are updated reliably and γ can be increased without instability. (vi) cmissc_miss inflates safety margins with age-of-information; larger values are beneficial under packet drops (more conservative motion), but become unnecessarily restrictive at high delays where the system is already sluggish, leading to a moderate co-design choice. V-B ROS2 QoS-based Reliability Analysis-Experiments A key parameter for practical experiments is the ROS 2 DDS middleware, which provides QoS policies to trade off delivery guarantees and latency [9]. In this evaluation, we analyze how QoC varies across different QoS settings, experimentally. We focus on the reliability policy, implemented at the application layer over user datagram protocol (UDP). Within this, BEST-EFFORT minimizes overheads by preceding retransmissions, prioritizing low latency while tolerating packet loss. Conversely, RELIABLE ensures delivery through application-layer retransmissions, resulting in latency characteristics distinct from transmission control protocol (TCP). Fig. 8, plots the QoC score, which is significantly lower under BEST-EFFORT QoS, as robots frequently miss timely updates, degrading cohesion. This is further reinforced by the delay-reliability relations, leading to higher packet deadline misses under BEST EFFORT QoS, as captured by the cumulative distribution function (CDF) and packet loss quantification in Fig. 9, which also shows the 50, 95, and 99 percentile delays for both QoSs. While RELIABLE QoS restores performance (51.5 % higher QoC), it introduces higher throughput variability (leading to higher 5G PRB demand if resources are dimensioned) and demand, as illustrated by the increased mean and standard deviation of throughput. This provides a critical trade-off for network providers, as it implies that a system optimized for high QoC may require more aggressive resource over-provisioning at the 5G base station. These results confirm that reliability choices directly shape the measured QoC and energy efficiency. Figure 8: Experimental observations on QoC variation and throughput for different QoS reliability. Figure 9: CDF of observed delay distributions and packet loss percentage for different QoS reliability. VI Conclusions and Future Work This work introduces a validated QoC-based abstraction designed to characterize the performance of collaborative robotic systems under stochastic network conditions. By incorporating non-holonomic kinematic constraints, we extend past QoC frameworks to bridge the gap between theoretical models and practical, edge-offloaded AGV navigation with dynamic collision avoidance. Towards this, we provide optimal operating points for various parameters for different delay-reliability points. Our experimental results, gathered over a private 5G testbed, closely align with simulation-based predictions, confirming the framework’s utility. Furthermore, a comparative analysis of ROS 2 transport policies demonstrates that RELIABLE QoS yields a 51.5% improvement in QoC over BEST-EFFORT, albeit at the expense of increased throughput variability. These findings provide a principled foundation for control-communication co-design in next-generation industrial automation. Future work will further calibrate the surrogate repulsion model against the Nav2/MPPI collision-avoidance stack by jointly varying costmap inflation, MPPI weights, and inter-robot spacing in larger experimental campaigns. VII Acknowledgements This work was supported in part by (1) Digital Futures project ‘Toward Safe Smart Construction: Algorithms, Digital Twins and Infrastructures’ under Grant VF 2020-0315, in part by (2) VR (Swedish Research Council) project ‘Optimal Sampling for Interactive Networked Applications’ under Grant 2022-03922 and lastly, (3) Styffes stiftelse. References [1] A. Baxi, M. Eisen, S. Sudhakaran, F. Oboril, G. S. Murthy, V. S. Mageshkumar, M. Paulitsch, and M. Huang (2022) Towards factory-scale edge robotic systems: challenges and research directions. IEEE Internet of Things Magazine 5 (3), p. 26–31. Cited by: §I. [2] G. Carabin, E. Wehrle, and R. Vidoni (2017) A review on energy-saving optimization methods for robotic and automatic systems. Robotics 6 (4), p. 39. Cited by: §I. [3] M. Groshev, L. Zanzi, C. Delgado, X. Li, A. d. l. Oliva, and X. Costa-Pérez (2025) Energy-aware joint orchestration of 5g and robots: experimental testbed and field validation. IEEE Transactions on Network and Service Management 22 (4), p. 3046–3059. External Links: Document Cited by: §I. [4] Z. Liu, J. Silva, R. Zhong, Q. Qin, N. Roy, V. Nan Fernandez-Ayala, J. Lesko, U. Håkansson, S. Sandberg, D. V. Dimarogonas, J. Gross, X. Vincent Wang, and L. Wang (2026) ConstrucTwin: digital twin-driven multirobot construction system toward industry 5.0. IEEE Transactions on Systems, Man, and Cybernetics: Systems (), p. 1–16. External Links: Document Cited by: §I. [5] S. Macenski, F. Martin, R. White, and J. Ginés Clavero (2020) The marathon 2: a navigation system. In 2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Cited by: §I-C. [6] S. Mohanti, D. Roy, M. Eisen, D. Cavalcanti, and K. Chowdhury (2023) L-norm: learning and network orchestration at the edge for robot connectivity and mobility in factory floor environments. IEEE Transactions on Mobile Computing 23 (4), p. 2898–2914. Cited by: §I. [7] S. Mostafavi, V. N. Moothedath, S. Ronngren, N. Roy, G. P. Sharma, S. Seo, M. O. Munoz, and J. Gross (2024) ExPECA: an experimental platform for trustworthy edge computing applications. In Proceedings of the Eighth ACM/IEEE Symposium on Edge Computing, SEC ’23, New York, NY, USA, p. 294–299. External Links: ISBN 9798400701238, Link, Document Cited by: 3rd item, §I-A, §IV-A. [8] M. Olguín Muñoz, N. Roy, and J. Gross (2022) CLEAVE: scalable and edge-native benchmarking of networked control systems. In Proceedings of the 5th International Workshop on Edge Systems, Analytics and Networking, EdgeSys ’22, New York, NY, USA, p. 37–42. External Links: ISBN 9781450392532, Document Cited by: §I. [9] H. Park, S. Lee, D. Um, H. Ryu, and K. Park (2025) An analytical latency model of the data distribution service in ros 2. In IEEE INFOCOM 2025 - IEEE Conference on Computer Communications, Vol. , p. 1–10. External Links: Document Cited by: §IV-A, §V-B. [10] K. Polachan, J. Pa, C. Singh, and T. Prabhakar (2022) Assessing quality of control in tactile cyber–physical systems. IEEE Transactions on Network and Service Management 19 (4), p. 5348–5365. Cited by: §I, §I. [11] Y. Qiao, Y. Fu, and M. Yuan (2023) Communication–control co-design in wireless networks: a cloud control agv example. IEEE Internet of Things Journal 10 (3), p. 2346–2359. External Links: Document Cited by: §I. [12] N. Roy, M. H. Dhullipalla, G. P. Sharma, D. V. Dimarogonas, and J. Gross (2025) Quality of control based resource dimensioning for collaborative edge robotics. In 2025 IEEE 22nd Consumer Communications & Networking Conference (CCNC), Vol. , p. 1–7. External Links: Document Cited by: 1st item, §I, §I, §I, §I-B1, §IV-C. [13] N. Roy, M. H. Dhullipalla, G. P. Sharma, S. Sandberg, D. V. Dimarogonas, and J. Gross (2026) Quality of control-based control-communication co-design for collaborative robotics. IEEE Transactions on Industrial Informatics 22 (6), p. 5219–5230. External Links: Document Cited by: 1st item, §I, §I, §I-B1, §I-B2, §I. [14] A. A. Sardar, A. S. Rao, T. Alpcan, G. Das, and M. Palaniswami (2024) Network resource allocation for industry 4.0 with delay and safety constraints. IEEE Transactions on Cognitive Communications and Networking 10 (1), p. 223–237. External Links: Document Cited by: §I. [15] A. Singletary, K. Klingebiel, J. Bourne, A. Browning, P. Tokumaru, and A. Ames (2021) Comparative analysis of control barrier functions and artificial potential fields for obstacle avoidance. In 2021 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), Vol. , p. 8129–8136. External Links: Document Cited by: §I-B2, §I-B2. [16] S. Tayade, P. Rost, A. Maeder, and H. D. Schotten (2020) Impact of short blocklength coding on stability of an agv control system in industry 4.0. In 2020 IEEE International Conference on Communications Workshops (ICC Workshops), Vol. , p. 1–6. External Links: Document Cited by: §I, §I-B1, §I-B1. [17] D. Urbaniak, S. Bro Damsgaard, W. Zhang, J. Rosell, R. Suárez, and M. Suppa (2024) Distributed control for collaborative robotic systems using 5g edge computing. IEEE Access 12 (), p. 148706–148718. External Links: Document Cited by: §I, §I. [18] G. Williams, P. Drews, B. Goldfain, J. M. Rehg, and E. A. Theodorou (2018) Information-theoretic model predictive control: theory and applications to autonomous driving. IEEE Transactions on Robotics 34 (6), p. 1603–1622. External Links: Document Cited by: §I-C. [19] D. T. Xuan, N. T. Hung, and V. T. Thang (2025) A comprehensive review of improved a* path planning algorithms and their hybrid integrations. Automation 6 (4). External Links: Link, ISSN 2673-4052, Document Cited by: §I-C.