In dynamic transportation environments, vehicular edge computing has emerged to address computation-intensive and latency-sensitive vehicular applications by executing computational tasks closer to moving vehicles at the network edge. When onboard vehicle hardware cannot process large workloads independently, transferring computational burdens to fixed infrastructure such as roadside units or remote cloud facilities often encounters transmission delays, backhaul bandwidth bottlenecks, or computational congestion. A research investigation led by Duan Xue, Gang Liu, Yulin Zhang, Bin Liu, Guohua Xu, and Xingying Huo, published in the Journal of King Saud University Computer and Information Sciences, evaluates an integrated framework that pairs roadside edge servers with opportunistic vehicles through digital twin technology and federated multi-agent reinforcement learning to optimize task completion delay, energy consumption, and task completion rates.
The study models modern vehicular environments where applications such as autonomous driving, augmented-reality navigation, cooperative perception, high-definition map updating, and vehicle-road safety services impose strict requirements on service continuity and real-time execution. While conventional vehicular edge computing deploys edge servers at fixed roadside units, the computational capacity and physical coverage of roadside units remain restricted, which risks queue delays and communication bottlenecks during dense traffic or bursty task arrivals. To expand computational availability beyond roadside units, the authors explore incorporating opportunistic vehicles that possess idle processor capacity and traveling trajectories amenable to temporary collaboration, allowing them to function as opportunistic edge servers through single-hop vehicle-to-vehicle communication links.
Relying on opportunistic vehicles introduces challenges because vehicle mobility, changing relative distances, fluctuating channel states, and uncertain link durations can cause task offloading failures if vehicles depart before computation or result transmission concludes. To mitigate these dynamic factors, the proposed architecture deploys a digital twin layer on cloud infrastructure. This digital twin constructs a virtual representation of the physical vehicular network by continuously synchronizing vehicle operational states, network topologies, task queues, wireless channel conditions, and edge resource distributions. By simulating network operations and predicting short-term resource availability, vehicle connection stability, and edge loads, the digital twin coordinates virtual information links that alleviate information lag and partial observability across the physical vehicular network.
The underlying network architecture is organized into a vehicle layer, an edge computing layer, and a digital twin layer. Within the vehicle layer, user vehicles continuously generate computational tasks while navigating the network, while opportunistic vehicles with idle resources serve as potential auxiliary compute nodes. The edge layer includes fixed edge servers mounted on roadside units alongside opportunistic edge servers formed by participating vehicles. The digital twin layer, deployed on cloud servers across backhaul connections, captures the global network view. To evaluate link viability, the authors formulate an opportunistic vehicle relationship model that determines social trust values between vehicles based on speed similarity and direction similarity, admitting only candidate vehicles exceeding a designated trust threshold into a dynamic trusted candidate set.
Under the formulated mathematical model, each user vehicle generates a computational task divisible into subtasks defined by data volume, required processor cycles, and a maximum tolerable delay deadline. Each subtask can be executed locally on the user vehicle, offloaded to a fixed edge server via vehicle-to-roadside wireless links, or offloaded to an opportunistic vehicle via vehicle-to-vehicle wireless links. Transmission delays, computing delays, energy consumption, server resource shares, and queue waiting times are mathematically formulated for each execution mode. A binary completion metric confirms success only when an entire task concludes within its maximum allowable delay without being truncated by vehicle departures out of communication range.
The joint optimization problem is formulated as a mixed-integer nonlinear programming model, classified as NP-hard, aimed at minimizing a weighted objective composed of total delay, total energy expenditure, and the reciprocal of the task completion rate. Constraints ensure that each subtask is mapped to exactly one processing location, allocated computing proportions remain within server capacities, task completion durations do not exceed specified deadlines, and computing frequencies stay within maximum hardware bounds. Weighting coefficients for delay, energy consumption, and task completion are established through the Analytic Hierarchy Process to systematically balance trade-offs among conflicting operational goals.
To solve the continuous action space optimization problem in a multi-vehicle environment, the researchers developed an algorithm combining multi-agent deep deterministic policy gradient techniques with federated learning. Within this framework, vehicle agents maintain actor networks that map local observations to continuous execution mode preferences and resource allocations, while critic networks leverage time-aligned global system states provided by the digital twin to evaluate policy performance. The critic network updates using target networks and double Q-learning principles through mean squared error minimization. The actor network updates parameters via deterministic policy gradient ascent, while target network parameters transition gradually using small tracking coefficients to ensure numerical stability during training.
Federated learning is integrated into the training procedure to address non-stationarity, local training bias, and client cold-start issues without requiring vehicles to share raw data. Vehicles train their local actor policies and periodically upload policy network weights to a cloud server every designated number of aggregation steps. The cloud executes a weighted parameter aggregation based on the capacity of each participant’s experience replay buffer and redistributes the global actor policy back to edge nodes. Meanwhile, critic networks can be updated centrally using global digital twin states or trained locally with synchronized state updates, maintaining consistent value assessments across agents while preserving vehicle privacy.
The authors evaluated the proposed FL-MADDPG method through numerical simulations using PyTorch coupled with the microscopic traffic simulator SUMO. The simulation modeled a 2.4 kilometer by 2.4 kilometer urban grid containing bidirectional roads with two lanes per direction, a 50 kilometer per hour speed limit, and vehicle arrivals governed by a Poisson process averaging approximately 2.2 vehicles per second. Over an 1800-second simulation period generating 3864 vehicle trips and 858,053 second-level records, average online vehicle presence was 477 vehicles with a peak of 550 vehicles. The communication setup comprised six uniformly spaced roadside units with 400-meter radio coverage radii, an orthogonal frequency-division multiple access channel with 20 sub-bands of 20 megahertz each, and digital twin state synchronization occurring at 3-second intervals.
Experimental performance was assessed against four baseline strategies: local computing, a fixed edge server only scheme, federated deep Q-networks, and conventional deep deterministic policy gradients. Across 6000 training episodes, the FL-MADDPG algorithm demonstrated smoother convergence and lower reward variance compared to baselines, which suffered from exploration instabilities or late-stage performance degradation. Under increasing counts of user vehicles and larger subtask data sizes, FL-MADDPG maintained lower average task response delays and superior task completion rates. Similarly, when examining edge server computing capacities, offloading to collaborative opportunistic vehicles relieved queue backlogs, whereas schemes relying solely on roadside servers suffered queue overflow and degraded service continuity.
Ablation experiments isolating specific system components showed that removing digital twin state predictions, federated policy aggregation, or opportunistic edge server resources significantly degraded performance. Excluding the digital twin removed global visibility from the critic network, excluding federated learning slowed model convergence and trapped decentralized agents in local optima, and excluding opportunistic vehicles escalated queuing delays during resource congestion. The authors conclude that integrating digital twin awareness with federated reinforcement learning effectively balances edge workloads, while noting future investigations will examine sixth-generation communication integration, lightweight digital twin modeling, and scalable critic architectures over broader vehicular mobility datasets.
In the mathematical formulation of execution modes, local energy consumption is calculated based on the CPU architecture of each vehicle, where energy per cycle scales with the square of the processor clock frequency due to its linear relationship with supply voltage. For offloading to roadside edge servers, transmission rates depend on channel bandwidth, transmission power, path loss exponents, background noise, and multi-vehicle interference across frequency-flat fading wireless channels. When tasks are routed instead to opportunistic edge servers, the total execution delay explicitly incorporates a queue waiting duration. This waiting interval is calculated by summing the required execution cycles of all preceding subtasks already queued at that specific opportunistic vehicle divided by its available computing frequency, capturing the cumulative response delay imposed by prior assignments.
To guide policy convergence across continuous action spaces, the reinforcement learning reward function integrates the negative exponential of overall system cost alongside an explicit trust bonus that rewards the selection of opportunistic vehicles with higher social trust scores. This formulation penalizes operational violations whenever assigned tasks breach resource sharing capacities, exceed individual delay deadlines, or surpass the maximum processor clock limits defined for user vehicles, roadside units, or opportunistic helper vehicles. Social trust values themselves combine direction similarity and speed similarity through weighted coefficients summing to one, assigning binary values based on directional alignment so that vehicles traveling in opposite directions cannot qualify as reliable relay nodes.
Computational complexity analysis indicates that local training scales with the number of user vehicle agents, mini-batch sample sizes, and the combined parameter updates of actor and critic networks. Because the centralized critic network processes global digital twin states and joint multi-vehicle actions, its computational demands exceed those of the decentralized actor network. Over complete training runs spanning multiple episodes, system complexity incorporates both local gradient evaluations and periodic federated aggregations, where the cloud server executes weighted parameter averaging based on replay buffer sizes without incurring the communication overhead of raw data exchanges.
Subject of Research: Digital twin-driven collaborative task offloading in vehicular edge computing networks
Article Title: Digital twin-driven collaborative task offloading method for vehicular edge computing
Article References: Xue, D., Liu, G., Zhang, Y., Liu, B., Xu, G., & Huo, X. (2026). Digital twin-driven collaborative task offloading method for vehicular edge computing. Journal of King Saud University Computer and Information Sciences, 38(9), Article 977. https://doi.org/10.1007/s44443-026-01410-1
Image Credits: AI Generated
DOI: 10.1007/s44443-026-01410-1
Keywords: Vehicular Edge Computing, Digital Twin, Federated Learning, MADDPG, Task Offloading, Roadside Units, Opportunistic Vehicles, V2V Communication, V2R Communication, SUMO Simulation, Edge Servers, Latency Optimization, Resource Allocation, Internet of Vehicles, Intelligent Transportation Systems, Mixed-Integer Nonlinear Programming
Cite Scienmag News
Marilyn Langley. (November 21, 2026). Digital Twin Framework Coordinates Vehicular Edge Computing with Opportunistic Vehicles. Scienmag. https://scienmag.com/digital-twin-framework-coordinates-vehicular-edge-computing-with-opportunistic-vehicles/
Marilyn Langley. "Digital Twin Framework Coordinates Vehicular Edge Computing with Opportunistic Vehicles." Scienmag, 21 November 2026, https://scienmag.com/digital-twin-framework-coordinates-vehicular-edge-computing-with-opportunistic-vehicles/. Accessed 21 November 2026.
Marilyn Langley. "Digital Twin Framework Coordinates Vehicular Edge Computing with Opportunistic Vehicles." Scienmag. November 21, 2026. https://scienmag.com/digital-twin-framework-coordinates-vehicular-edge-computing-with-opportunistic-vehicles/








