Optimized multi-tier task offloading strategy for sustainable IoV systems in 6G networks

Abstract

Sixth-generation (6G) networks are likely to support advanced Internet of Vehicles (IoV) applications that have rigid latency, reliability, and computation demands. Nevertheless, efficient task offloading is a challenging problem because vehicle environments are characterized by mobility, changing channels, varying task requirements, constrained edge resources, and growing energy demands. To address these issues, this study presents an Optimized Multi-Tier Task Offloading Strategy (OMTOS) for sustainable IoV systems. The proposed framework comprises a four-tier computing architecture comprising vehicles, roadside units (RSUs), mobile edge computing (MEC) servers, and cloud infrastructure. The generalized latency-energy optimization problem is formulated to allocate tasks across these levels, accounting for task due dates, resource capacity, communication delay, computation delay, and energy consumption. To address dynamic offloading, OMTOS employs a centralized training and decentralized execution (CTDE) based multi-agent Soft Actor-Critic (SAC) method, where the vehicle agents can make decentralized offloading decisions with centralized critics guiding the coordinated learning process during training. It is tested against rule-based and heuristic as well as deep reinforcement learning and various multi-agent reinforcement learning baselines, including LE, EO, RO, GO, DQN, DDPG, SAC, MADDPG, and MAPPO. The aforementioned results reveal that OMTOS achieves low average delay, low energy consumption, a high task success rate, and high convergence compared to the competing methods. Sensitivity analysis also indicates that the latency and energy weightings can be changed to suit various IoV service requirements, including delay-critical safety services, and energy-conscious delay-tolerant services. These results show that OMTOS offers an adaptive and sustainable task-offloading tool in 6G-enabled IoV environments.

Introduction

The development of beyond fifth-generation (B5G) and sixth-generation (6G) networks is transforming the Internet of Vehicles (IoV) through ultra-reliable and low-latency communication and scaled connectivity. Potential applications of modern vehicles, such as autonomous driving, smart traffic systems, and even safety services, generate real-time, delay-sensitive, and computationally intensive tasks^1,2. Applications are sensitive to speed and decision-making, and therefore cannot be effectively handled by traditional cloud-centric architectures because transmission delays are high and the network is congested^3,4. This has led to task offloading being considered as a basic process in the IoV system, where vehicles can offload computation tasks onto external resources.

Mobile Edge Computing (MEC) has significantly enhanced system responsiveness by placing computing resources closer to vehicles. The current offloading processes, however, are based on single- or two-tier architectures, where either tasks are run locally or offloaded to edge servers^5–7. These methods cannot be used in highly dynamic IoV environments that involve vehicle mobility, varying network conditions, task heterogeneity, and limited edge resources. These limitations are the ones that are prone to imbalance in resource, increase the queueing time and poor use of the computing infrastructure^8–10. To beat these obstacles, multi-tier task offloading has emerged as a tempting paradigm. Services can be allocated in a relaxed manner across a multi-layer architectural framework comprising multiple computation layers, such as vehicles, roadside units (RSUs), MEC servers, and cloud infrastructure^11,12. This would make it possible to allocate tasks with greater flexibility and in a smarter way, with delay-sensitive tasks performed at lower levels (e.g., vehicles or RSUs) and high-processing-intensity tasks offloaded to higher levels with more processing resources. This flexibility is critical, especially in 6G-based IoV systems, where bespoke communication and computing are anticipated to enhance highly dynamic, context-based services.

Besides performance efficiency, sustainability has also been a serious issue in next-generation vehicular networks. The rapid development of IoV systems, the constant exchange of data, and the processing of data are significant factors in increasing energy consumption, both on vehicles and at the edge of the infrastructure^13,14. Although most existing studies have focused on reducing latency, it often fails to elaborate on the significance of operational efficiency for energy utilization. But to achieve sustainable IoV systems, there must be concessions in latency and energy use to achieve concrete performance and environmental efficiency^15,16. Thus, it is necessary to study energy-consciousness in the task redistribution during the design of green, scalable IoT systems for 6G networks. However, with the current developments, there are still some gaps in study. First, the current literature is mostly limited to simplified models and does not fully exploit the potential of multi-tier collaborative computing. Second, the heterogeneity in task size, computational demands, and delay sensitivity is not adequately accounted for in most cases¹⁷. Third, the concept of sustainability, especially the idea of offloading that is energy-efficient at various levels, remains scarce. Lastly, the changing characteristics of IoV environments require more flexible and effective policies that enable efficient decision-making across different system conditions^18,19.

To address these challenges, this paper proposes an Optimized Multi-Tier Task Offloading Strategy (OMTOS) for sustainable 6G-enabled IoV systems. In contrast with traditional solutions that focus on the elements of latency-sensitive offloading, energy-sensitive optimization, MDP-based decision-making, or resource allocation based on reinforcement learning, OMTOS incorporates all the elements of latency-sensitive offloading, energy-sensitive optimization, MDP-based decision-making, or resource allocation based on reinforcement learning into the unified multi-layer framework. The suggested strategy is dynamic because it assigns heterogeneous vehicular tasks to vehicles, RSUs, MEC servers, and cloud resources, taking into account procedures for latency, energy, task deadline, resource availability, mobility, and the wireless channel. Through the exploitation of four computation tiers together with the adoption of a CTDE-informed multi-agent SAC mechanism, OMTOS allows the coordinated learning during training and the decentralized offloading decisions during execution, offering an application-aware and flexible solution to the challenge of a dynamic IoV environment. The key findings of this work can be summarized as follows:

• This paper proposes an IoT task-offloading framework based on a four-layer network architecture comprising vehicles, RSUs, MEC servers, and cloud infrastructure. This framework enhances task allocation flexibility more than traditional single-tier or two-tier offloading models.

• This work develops a generalized latency-energy optimization model for heterogeneous vehicular tasks. The work considered the task data sizes, task CPU-cycle demand, task delay tolerance, communication delay, computation delay, resource capacity, and energy consumption.

• The proposed framework considers mobility-awareness in the formulation of task offloading by incorporating time-varying channel conditions, vehicle location, RSU association, and handover delay into the system model and decision-making.

• The proposed OMTOS algorithm will utilize a CTDE-driven multi-agent SAC architecture with decentralized vehicle-side actor policies trained by a single centralized critic.

• To prove the efficiency of OMTOS, the study will evaluate the proposed framework with rule-based, heuristic, DRL, and MARL baselines, such as LE, EO, RO, GO, DQN, DDPG, SAC, MADDPG, and MAPPO. The outcomes include better average delays, energy usage, task success rates, convergence behavior, scalability, and adaptability to varying task sizes.

• A sensitivity analysis of the latency-energy weighting factors is also carried out in this study to demonstrate that OMTOS can be tailored to meet the various manufacturing to production process requirements, and the manufacturing to production process needs, of the requirements of various types of IoV services, including, but not limited to, delay-oriented manufacturing to production process needs, energy-aware delay-tolerant services.

The rest of this paper is structured as follows. Section “Related Work” discusses pertinent literature about task offloading and IoV systems. The system model is presented in Section “System Model”. The optimization problem is formulated in Section “Problem Formulation”. Section "Proposed Optimized Multi-Tier Task Offloading Strategy" outlines the proposed multi-tiered offloading strategy. Section "Simulation Setup and Performance Evaluation" presents the simulation setup and performance evaluation, and the paper concludes in Section "Conclusion and Future Work" with future directions. The main abbreviations used throughout the manuscript are summarized in Table 1.

Table 1.

List of abbreviations used in this study.

Abbreviation	Full Form	Abbreviation	Full Form
6G	Sixth-generation	B5G	Beyond fifth-generation
IoV	Internet of Vehicles	OMTOS	Optimized Multi-Tier Task Offloading Strategy
RSU	Roadside Unit	RSUs	Roadside Units
MEC	Mobile Edge Computing	VEC	Vehicular Edge Computing
MEC server	Mobile Edge Computing server	RL	Reinforcement Learning
DRL	Deep Reinforcement Learning	MARL	Multi-Agent Reinforcement Learning
SAC	Soft Actor-Critic	MASAC	Multi-Agent Soft Actor-Critic
CTDE	Centralized Training and Decentralized Execution	MDP	Markov Decision Process
CSI	Channel State Information	QoS	Quality of Service
DQN	Deep Q-Network	DDPG	Deep Deterministic Policy Gradient
MADDPG	Multi-Agent Deep Deterministic Policy Gradient	MAPPO	Multi-Agent Proximal Policy Optimization
LE	Local Execution	EO	Edge-Only Offloading
RO	Random Offloading	GO	Greedy Offloading
NOMA	Non-Orthogonal Multiple Access	UAV	Unmanned Aerial Vehicle

Related work

In the IoV, task offloading has attracted significant interest due to the rapid proliferation of latency-sensitive and computationally intensive vehicle applications. In^20,21, MEC was introduced into the vehicular network at an early stage to consider the issues of latency reduction and responsiveness enhancement. Similarly, the report in²² explores binary offloading schemes, in which tasks are either executed locally or offloaded to local edge servers, while addressing trade-offs among delay reduction, scalability, and resource consumption. In addition, recent studies^23–25 suggest partial offloading schemes that distribute tasks across multiple nodes, thereby enhancing flexibility, but these schemes introduce greater complexity in scheduling and coordination. Most such models, as explained in^26,27, assume static or simplified network conditions, which makes them restrictive in the highly dynamic IoV environment with mobility and varying communication links.

In^28,29, the authors present multi-tier and hierarchical computing networks to address the constraints of single-layer offloading. These architectures include numerous layers, i.e., vehicles, RSUs, edge servers, and cloud platforms to provide greater flexibility in the task distribution. Moreover, the study by^30,31 shows that multi-tier systems can significantly improve system performance through heterogeneous computing resources. However, as described in³², the majority of current models have covered only two- or three-level models and do not fully utilize the entire hierarchy of the IoV. Also, the study in^33,34 indicates that communication across layers is typically suboptimal, which contributes to higher overhead and inefficient task distribution. It therefore follows that the research problem of integrating decision-making on all levels is open.

Various optimization methods have been studied to address the complexity of decision-making in task offloading. In^35,36, heuristics and convex optimization are used to find efficient ways to offload when resources are constrained. Similarly, game-theoretic solutions in³⁷ equilibrate the interactions among network entities. However, these solutions are not always flexible in dynamic IoV environments. In recent years, the machine learning-based methods have become popular. Deep Q-Networks (DQN) are employed for offloading policy learners in^38,39, and Deep Deterministic Policy Gradient (DDPG) is applied in continuous decision space⁴⁰. In addition, the experiment in⁴¹ uses Soft Actor-Critic (SAC) to improve stability and convergence in complex environments. Moreover, the methods of multi-agent reinforcement learning (MARL) in⁴² enable distributed nodes to exchange information and make coordinated decisions. Even after these developments, the majority of existing learning-based methods are primarily concerned with minimizing latency or maximizing throughput, as explored in⁴³, and do not prioritize energy efficiency or sustainability. Besides, many approaches rely on centralized training or information at a global information, which is not scalable for large-scale IoV deployments.

The optimization of network scheduling during computing power and optimization of 6G edge computing further added to the research background of intelligent allocation of tasks. In⁴⁴, computation-power netbox scheduling is explored to coordinate distributed computing resources and communication links, demonstrating how computation-aware scheduling can optimize the utilization and responsiveness of network computing resources and links in networked edge computing. Similarly⁴⁵, discusses 6G-aware edge computing optimization, in which communication, computing, and service demands are comprehensively addressed to support delay-sensitive and resource-intensive applications. These researchers determine the significance of integrated computing resources in the future intelligent networks. However, they mainly focus on general computing power network scheduling or 6G edge optimization, but not specifically in multi-tier task offloading in IoV systems through the joint consideration of vehicles, RSUs, MEC servers, and cloud infrastructure. Besides, the proposed OMTOS scheme combines latency-energy optimization with a multi-agent SAC-based learning plan to support adaptive, sustainable solutions for offloading choices under dynamic vehicle conditions.

Sustainability has also become an essential demand for next-generation IoV systems. The authors in^46,47 include energy consumption in task offloading by optimizing transmission power and computation resources. In the same fashion, recent studies^48,49 aim at minimizing the energy footprint of a vehicular network at the expense of still ensuring satisfactory levels of performance. However, as the authors of^50,51 point out, most energy-aware solutions are constrained by a one-tier or simplified architecture and do not address the complexity of multi-tier IoV systems. In addition, communication between the various layers of the network and its effect on energy efficiency are not well studied. Thus, the increasing demand is to develop a complex set of structures, which jointly optimize latency and energy consumption and exploit multi-tier structures.

The study in⁵² introduced present a framework in which implemented a real-world UAV delivery system by integrating MEC and blockchain-based resource management. Based on this work, it is possible to support practical computation, communication, and data management processes in aerial edge environments using UAV-edge prototypes. In^53–55, the authors further examined a UAV-based MEC system that dynamically provides services, performs margin processing onboard, and provides intelligent application support. In addition, these field tests affirm the validity of UAV-assisted edge computing for future external vehicular and edge network services. However, in [3], UAV-mounted MEC nodes are not covered by the present OMTOS system model, which focuses on ground-based IoV task offloading among vehicles, RSUs, MEC servers, and cloud infrastructure. Thus, extending OMTOS to UAV-enabled MECs is a promising research direction for the future.

Recent work has tried to enhance the offloading of vehicle tasks through mobility awareness, decision-making using DRL, and coordinating tasks among agents⁵⁶. examined a vehicle edge computing task-offloading strategy with the use of reinforcement learning, and vehicle position, velocity, and computing resources were among the aspects that were taken into account. They propose a V-DQN-based technique that enhances service stability but focuses more on vehicle-edge resource selection and does not fully leverage the four-tier vehicle-RSU-MEC-cloud hierarchy for joint latency-energy optimization. Similarly^57,58, suggested a platform hybrid hierarchical DRL agent to complete task offloading and resource allocation in intelligent connected vehicles. Their method breaks down the decision process into an upper DDQN layer that communicates decisions and a lower DDPG layer that decides on power control, offloading ratio, and resource allocation. Despite the fact that this decomposition enhances the adaptability in dynamic environments, the framework still is centered on vehicle–road collaborative MEC and does not explicitly model cloud-tier participation or CTDE-based multi-agent SAC coordination.

Recent multi-agent schemes have gone further to enhance the coordination between the vehicles and edge nodes⁵⁹. have put forward a multi-agent deep reinforcement learning policy for vehicular edge computing using NOMA, queuing theory, and a stochastic game formulation to optimize delay, energy consumption, cost, and server utilization. The approach, however, predominantly focuses on multi-agent edge offloading and is not developed as a sustainability-oriented multi-layered offloading approach between vehicles, RSUs, MEC servers, and cloud infrastructure^60,61. suggested an optimization process based on cooperation between vehicle agents and RSU agents when there are resource constraints, and centralized coordination in RSU agents is necessary. This approach enhances task completion rate, delay, and energy performance; however, it remains focused on task coordination at the RSU level and does not directly address the broader multi-tier latency-energy trade-off introduced in the present work.

Emergent AI-assisted offloading systems and digital twins have also been researched^62,63. proposed a VEC framework based on digital replicas to assist decisions made by multiple tasks generated by vehicles in a single time slot. This enhances monitoring and task control, but does not focus on sustainable multi-tier offloading using CTDE-based multi-agent SAC. More recently, it has been investigated to enhance adaptive, low-latency decision-making at RSU edge nodes with the assistance of LLMs. These approaches are promising to high-dimensional reasoning and real-time adaptation, although their practical implementation can raise other model-SIZE, inference-latency, and RSU computational-overhead issues.

Even though recent research on the topics of offloading of tasks to vehicular computers, scheduling of computing resources, edge-assisted vehicular computing, and reinforcement-learning-based resource allocation have made significant strides, there are still several limitations in this area of research. Most current techniques still rely on local-edge, RSU/MEC-based, single-tier, or limited two-tier offloading models and fail to fully leverage a four-tier hierarchy that includes vehicles, RSUs, MEC servers, and cloud infrastructure. Moreover, latency, energy consumption, mobility, channel variation, task heterogeneity, and resource availability are commonly studied independently, which restricts the realism and flexibility of offloading decision-making in dynamic multi-user IoV surgery applications. Certain DRL implementing approaches exploit single-agent, centralized, or weakly coordinated learning, and it is hard to obtain resource contention among multiple cars. To fill these gaps, OMTOS develops sustainable task offloading across vehicles, RSUs, MEC servers, and cloud resources, taking into account both latency and energy, as well as task due dates, mobility-aware channel variation, handover effects, and the availability of computation resources. By further extending the CTDE-based multi-agent SAC design to enable decentralized vehicle-side decisions with centralized critic-based training, OMTOS will be able to deploy a universal 6G-enabled IoV system that integrates multi-tier flexibility, latency-energy consciousness, mobility-conscious modeling, and coordinated multi-agent learning in a single 6G-enabled IoV platform.

System model

This section provides the system architecture, task model, communication model, computation model, and energy model of the proposed multi-tier task offloading in a 6G-enabled IoV environment. Fig. 1 shows a multi-tier architecture for the OMTOS and demonstrates how the vehicles, RSUs, MEC servers, and cloud infrastructure interact within the proposed IoV offloading structure. This figure also presents how computation tasks arising on vehicles can be distributed to lower levels of computation, taking into account available resources, timing constraints, and network states. It is important to note that the current framework is based on a UAV-assisted computing node rather than a ground-based four-tier IoV architecture.

Fig. 1.

Multi-tier architecture of the proposed OMTOS framework. Vehicles generate heterogeneous computation tasks that may be executed locally or offloaded to RSUs, MEC servers, or cloud infrastructure, depending on task requirements, communication conditions, and available computing resources.

We base our discussion around a multi-tier IoV system that is made up of four levels, such as the vehicle layer (Tier 1), roadside unit (RSU) layer (Tier 2), MEC layer (Tier 3), and the cloud layer (Tier 4). This hierarchical structure enables flexible task execution across the various levels of computation, enabling the system to effectively use distributed resources and adapt to different network conditions. Let denote the set of vehicles, represent the set of RSUs, and denote the set of MEC servers. Each vehicle in a system type performs intensive computations that may be distributed to lower levels or offloaded to higher levels when resources are available, latency requirements can be met, and the network conditions allow. To formulate the offloading decision, a binary variable is introduced to indicate whether the task generated by vehicle i is allocated to tier k. The offloading decision variable is given in Eq. (1):

1

where corresponds to local execution at the vehicle, RSU processing, MEC server execution, and cloud processing, respectively. To ensure that each task is assigned to exactly one processing location, the constraint in Eq. (2) is imposed:

2

This constraint ensures that no job is designed in the system beyond the chosen level, thereby maintaining stable task assignment and eliminating redundant processing. The vehicles generate a computation task with a data size, a computation requirement, and a delay limit. The formalization of the task is presented in the form of Eq. (3):

3

where is the size of the input data in bits, is the number of CPU cycles required to execute the task, and is the maximum delay that the task can allow. These parameters capture the heterogeneity of vehicular applications and their usage, enabling the system to make adaptive offloading decisions based on each task’s demand and the limitations on quality of service (QoS). The communication model shows how data is passed between vehicles and other computational levels within the IoV system. Shannon capacity formula takes the following form of a triangular value of the transmission rate between vehicle i and node j in the form of the following Eq. (4):

4

in which B is the bandwidth available, is the power used by vehicle i to transmit power or signal to node is the gain of the channel between vehicle i and node j and is the noise power. The Formula above summarizes the impact of wireless channel conditions on the achievable data rate.

Since IoV nodes are mobile, the communication channel between vehicles and computation tiers varies over time. Let denote the position of vehicle i at time t, denote its velocity vector, Δ t denote the scheduling interval, and denote the location of the computation-tier node associated with tier k. The mobility-aware position update and the distance between vehicle i and tier k are defined in Eqs. (5) and (6), respectively:

5

6

Based on the distance-dependent mobility model, the channel gain and achievable transmission rate are computed using Eqs. (7) and (8), respectively:

7

8

where is the small-scale fading component, ν is the path-loss exponent, B is the bandwidth, is the transmission power of vehicle i, and is the noise power. This expression can be used to set the channel condition and the transmission rate to depend on vehicle movement rather than a fixed deterministic channel.

In RSU-assisted offloading, the strong channel available is used to pick the serving RSU, and the indicator to signal the handover is defined as whether the serving RSU transitions between two consecutive time steps. These two operations are expressed in Eqs. (9) and (10), respectively:

9

10

Using the mobility-aware rate and handover indicator, the service delay is formulated in Eq. (11):

11

where is the task data size, is the required CPU cycles, is the computing capacity allocated to task i at tier k, denotes the handover-related signaling and service-migration delay, and for local execution while for offloaded execution. By incorporating mobility-aware channel variation, RSU association, and handover delay, the proposed model better represents multi-user IoV offloading under dynamic vehicular movement. Eq. (12) gives the delay schedule for attempting to transmit a duty to tier k as a function of the communication rate.

12

where represents the dimension of the input data of the job, and the pace at which vehicle i forwards data to the tier picked k. This ratio measures the time required to transfer task information over the wireless channel and is of utmost importance to the total service delay. The computation model describes the delay incurred in the way activities are performed in the various levels of the IoV system. The computation delay at tier k is defined as shown in Eq. (13):

13

where denotes the required CPU cycles of task i, and represents the computing capacity of tier k. This formulation reflects the time required to process a task based on the available computational resources at the selected tier. The overall service delay for task execution at tier k includes both communication and computation delays, and is given in Eq. (14):

14

where represents the transmission delay and denotes the computation delay. For local execution at the vehicle, no transmission delay is incurred. Therefore, the total delay is solely determined by the local computation capability and is expressed in Eq. (15):

15

where denotes the local computing capacity of vehicle i. The energy consumption model accounts for transmission energy and the energy required to perform tasks in the IoV system. In the case of offloading to a higher tier, the major energy consumers are data transmission; in local execution, it is computing. The transmission energy required to offload a task from vehicle i to tier k is defined as given in Eq. (16):

16

where represents the transmission power of vehicle i, and denotes the transmission delay. For local task execution, the computation energy consumption is modeled as shown in Eq. (17):

17

where κ is the effective switched capacitance coefficient, denotes the required CPU cycles, and represents the local computing frequency of vehicle i. This formulation captures the relationship between computational load and energy consumption. The total energy consumption for executing a task at tier k is given in Eq. (18):

18

The reason for this definition is that local execution incurs computation energy, whereas offloading to higher tiers incurs transmission energy. The aim of the proposed framework is to reduce the total latency and energy consumption of the multi-tier IoV system. This is formulated as an optimization problem in which the offloading decisions are determined to achieve an optimal trade-off between delay and energy efficiency. The objective function is expressed as given in Eq. (19):

19

in which α and β are weighting values which regulate the proportional values of latency and energy consumption, respectively. It is possible to adjust these parameters so that, should there be demand on the applications, the system can provide delay-sensitive or power-efficient performance. The multi-tier IoV system has an optimization problem with a number of constraints to regulate task distribution, delay problems and computational capacity and also decision variables to maximize its feasibility and viability.

20

21

22

23

Under Eq. (20), all the tasks generated at any vehicle are broadcast to the respective tier of computation, eliminating redundancy. Eq. (21) uses delay as a requirement and thus will ensure that the time taken to complete each process does not exceed the maximum limit. The number of works per tier is constrained by allocating each tier an allowance using Eq. (22) to ensure it fits within processing. Finally, an introductory decision is made in Eq. (23) by assigning each task to either a specific tier or not. All these constraints are system-feasible and allow effective, realistic task offloading decisions.

The section presents an elaborate development of a multi-tier IoV system model that considers hierarchical computation at various levels, the heterogeneity of task features, the dynamic communication environment, and energy usage. In addition, a collaborative optimization model is formulated to achieve a balance between latency and energy usage, enabling the explanation of efficient and sustainable decisions related to task offloading within the system.

Problem formulation

In this section, we develop the multi-tier task offloading problem as a joint optimization problem and model it as a Markov Decision Process (MDP) to facilitate intelligent, adaptive decision-making in dynamic IoV settings. Based on the elaborated system model, the problem of achieving overall agreement on the total service delay and the energy capacity deployed by all vehicles and computation levels will arise. This is developed as a constrained optimization model in which the offloading decisions are determined to achieve an efficient trade-off between latency and energy consumption.

The multi-tier task-offloading problem is expressed using a unified tier-based notation. Let denote the set of computation tiers, where k=0 represents local vehicle execution, k=1 represents RSU execution, k=2 represents MEC execution, and k=3 represents cloud execution. The optimization is formulated as shown in Eq. (24):

24

where denotes the binary offloading decision variable, denotes the total service delay of task i when executed at tier k, and denotes the corresponding energy consumption. The parameters α and β are non-negative weighting factors that control the relative importance of delay and energy consumption, respectively. By adjusting these weights, the model can support a range of IoV service requirements, from delay-critical applications to energy-aware, delay-tolerant services. The optimization problem is subject to the following constraints:

25

26

27

28

The constraint in Eq. (25) ensures unique task allocation since each task produced by a vehicle is allocated to only one tier of calculation. The constraint of the deadline in Eq. (26) permits that the selected offloading decision would satisfy the maximum tolerable delay of the task constraint. The capacity limit in Eq. (27) prevents the workload scheduled to each level of computation from exceeding its processing capability within the time interval Δ t, where represents the available processing capability of level k. Lastly, Eq. (28) retains the binary representation of the offloading decision variable.

These constraints make sure that the task assignment is feasible, assigning each task to be scheduled at only one computation tier, without violation of the delay requirement, the processing capacity of each computation tier, and the binary nature of the offloading decision variables. These limitations, combined, make it possible to make realistic and effective task-offloading decisions in the proposed multi-tier IoV system.

The optimization model developed is extremely complicated and belongs to the class of mixed-integer nonlinear programs (MINLP), which are known to be NP-hard. Such binary decision variables, nonlinear objective functions, and constraints significantly complicate the process, making it difficult to find optimal solutions using conventional optimization techniques. Moreover, the problem is highly dynamic because it is affected by vehicle motion, time-varying channel conditions, and stochastic task arrival. These ambiguities further destabilize the system and worsen the decision-making process. In this case, smart, adaptable solutions to problems are needed, since conventional optimization techniques cannot address this predicament in real-time IoV.

This contribution presents the task offloading problem as a Markov Decision Process (MDP) to account for the dynamic and stochastic nature of IoV systems. This formulation helps the system arrive at the most appropriate decision-making policies by interacting with the environment over time. To capture the dynamic and stochastic nature of the proposed task-offloading problem, the optimization problem is reformulated as a Markov decision process (MDP). Formally, the MDP is represented by the tuple in Eq. (29):

29

where denotes the state space, denotes the action space, denotes the state-transition probability function, denotes the immediate reward function, and is the discount factor. This notation clearly distinguishes the MDP components from the previously defined sets of vehicles, RSUs, MEC servers, transmission power, and data-rate variables, thereby improving readability and consistency. The state space represents the dynamic network conditions, computing-resource availability, queue status, and task information in the IoV system. The system state at time t, denoted by , is defined as shown in Eq. (30):

30

where represents the queue status of the computation tiers, denotes the wireless channel conditions between vehicles and computation tiers, represents the available computing resources, and captures task-related information. For vehicle-level decision-making, the state can be represented more specifically as shown in Eq. (31):

31

where denotes the queue length associated with vehicle i, represents the channel gain between vehicle i and computation tier k, indicates the available computational capacity at tier k, denotes the task data size, represents the required CPU cycles of the task, and denotes the task deadline. This state representation enables the learning agent to adaptively make offloading decisions across variable network, computational, and task conditions.

The action space is the space of the possible offloading decisions at each time step. In the framework proposed, the action defines the level of computation to use when executing each task created by the vehicles. The overall action at time t is represented as shown in Eq. (32):

32

where denotes the offloading decision variable for vehicle i at tier k. For each vehicle i, the individual action is defined as shown in Eq. (33):

33

where 0, 1, 2, and 3 correspond to local execution, RSU execution, MEC execution, and cloud execution, respectively. Such a discrete action space helps the agent dynamically choose the most appropriate computation tier based on the current system state. The state-transition model provides a representation of the changes in the environment of the IoV when an offloading action is carried out. The transition dynamics are represented as shown in Eq. (34):

34

where denotes the probability distribution over the next state after taking action in state . This formulation captures the stochastic changes caused by vehicle mobility, task arrivals, wireless channel variations, queue updates, and available computing resources. These restrictions guarantee the task assignment to be feasible by assigning each task to a single computation tier, meeting the delay requirement, respecting the processing capacity of each computation tier, and preserving the binary nature of the offloading decision variables. Collectively, these constraints facilitate realistic and effective task-offloading decisions within the proposed multi-tiered IoV system.

This operation models the stochasticity of the IoV environment by accounting for changes in the system state resulting from offloading decisions. Some of the dynamics affecting state transitions are vehicle mobility, which affects network connectivity and topology; the influx of new computation tasks; changes in wireless channel conditions; and changes in queue lengths at various computation levels. All of them, collectively, lead to the ambiguity and complexity of the system dynamics, which makes the decision-making process very adaptive. The reward function is also designed to discourage increases in delay and energy consumption to jointly optimize latency and energy consumption. The reward at time t is defined as given in Eq. (35):

35

where α and β are weighting factors that balance the importance of latency and energy consumption. The negative sign is necessary to ensure that the reduction of delay and energy establishes a positive relationship between the accumulation of a reward and the purpose of the optimization problem. In order to deal with the violation of constraints, especially delay constraints, we can add to the reward by the inclusion of a penalty term. The modified reward formulation is given in Eq. (36):

36

where λ is a penalty coefficient that controls the severity of constraint violations. This formulation penalizes any delay beyond the maximum tolerable level, thus promoting the performance of the learning agent to meet the QoS stipulations and maximize performance. The learning framework aims at identify an optimal policy that maximizes the expected cumulative reward over time. This policy will direct the agent to make the best decisions for offloading, given the system’s changing conditions. The optimal policy is defined as given in Eq. (37):

37

where is the discount factor that determines the relative importance of immediate and future rewards, and denotes the expected value. This formulation ensures that the agent acquires a policy that maximizes short-term performance while also considering long-term system efficiency.

The proposed framework views the IoV system as a multi-agent environment in which all vehicles are independent agents. These agents share a common environment, that is, the IoV network, and make decentralized solutions whilst contributing to a shared system goal. Both agents monitor their local state and choose an action based on their decision to offload tasks. The agents are made to work together through a collective form of rewards or a global maximization objective, which promotes joint actions to enhance overall system performance. The joint action of all agents at time t is defined as given in Eq. (38):

38

where denotes the action taken by agent i. This joint action representation describes the common decision-making process with a single representative for all vehicles within the network, and the system can model interactions and dependencies among two or more agents.

The task offloading problem is formulated in this framework as a joint optimization problem to minimize latency and energy consumption. As a result of the combinatorial structure of the offloading choices and interaction between the agents, a computational problem is NP-hard. To efficiently address this complexity, the problem is modeled as a Markov Decision Process (MDP), with learning-based solutions incorporated to account for the dynamic and stochastic nature of the IoV environment. The state space, action space, reward function, and policy objective are the main elements of the MDP that are established to support intelligent, adaptable decision-making.

Proposed optimized multi-tier task offloading strategy

This section introduces the Optimized Multi-Tier Task Offloading Strategy (OMTOS) that is to be proposed to achieve sustainable 6G network IoV systems. The strategy will equally distribute tasks with smartness at various levels, as well as optimizing jointly on latency and energy usage within a dynamic environment. In contrast to more traditional independent offloading models, OMTOS considers the problem of determining the tasks performed by vehicles as a coordinated multi-agent, decision-making problem whereby local vehicle behavior modulates shared RSU, MEC, and cloud resources.

The offloading task problem stated in Section “Problem Formulation” is highly dynamic and computationally intensive, and cannot be solved using traditional optimization methods. To overcome this obstacle, a proposed solution is a reinforcement-learning-based approach. This solution enables the system to learn optimal offloading choices within a continuous interaction framework with the environment. In particular, the use of a SAC-based multi-agent architecture is expected to stabilize the learning process, effectively explore, and achieve successful performance in complex IoV environments. The method will enable various agents to make task offloading decisions jointly and adapt to time-delaying network conditions. The solution proposed employs multivehicle, roadside, and MEC servers, and cloud computing, which allows flexible allocation of tasks by covering all vehicles, the RSUs, and the MEC servers, in addition to the cloud infrastructure. Moreover, it coexists with the optimization of latency and energy usage, thereby providing a performance-sustainability balance. The architecture will fit dynamic vehicular environments (in terms of mobility and communication conditions) and provide scalable, distributed learning for a large-scale IoV system.

Under the proposed framework, a vehicle is represented as a separate agent that decides which tasks to offload based on its immediate predictions. Such a decentralized environment allows agents to work autonomously and to fail to achieve the system goal. A centralized training and decentralized execution (CTDE), paradigm is followed to coordinate the learning of numerous agents whereby agents are able to exchange information during training and independently make choices during execution. Fig. 2 shows the logical architecture and interaction flow of the proposed OMTOS framework under a centralized training and decentralized execution paradigm. The operationalization of the IoV environment at each decision step is in the form of state observations of queue conditions, channel quality, available computing resources, and task characteristics. Based on these observations, each vehicle agent decides where to take action during offloading within the actor network, either locally or offloaded to the RSU, MEC, or cloud tier. The identified actions are then implemented in the IoV environment, initiating the next state and a reward based on the joint latency-energy goal. The transitions observed during training are stored in the replay buffer and used to update the critic and actor networks during SAC learning. The architecture enables adaptive decisions about task offloading without sacrificing coordination when learning and performing lightweight tasks at the vehicle end.

Fig. 2.

Flow diagram of the proposed OMTOS framework showing the main agent, environment, action, reward, and learning-update.

Let N denote the number of vehicle agents. In the proposed CTDE-based multi-agent SAC formulation, each vehicle i observes a local observation , which may include its local queue length, channel condition, task size, CPU-cycle demand, delay tolerance, and locally available resource information. The global system state at time t is represented in Eq. (39):

39

where denotes the global computing-resource state of the RSU, MEC, and cloud tiers. Each vehicle agent uses a decentralized actor policy to select its own offloading action, as shown in Eq. (40):

40

where represents local execution, RSU execution, MEC execution, and cloud execution, respectively. The joint action of all vehicle agents is defined in Eq. (41):

41

In contrast to independent SAC, where the central critic learns about the local critic only through its hidden states, the proposed CTDE-based MASAC framework uses centralized critics during training. The centralized critic considers the world situation and the collective action of all agents, as it is described in Eq. (42):

42

where two critics are used to reduce overestimation bias, following the clipped double-Q mechanism of SAC. During execution, the vehicle agents do not require centralized critics; each vehicle uses only its local actor policy to make decentralized offloading decisions. The global reward is designed to reflect the overall latency–energy objective of the multi-tier IoV system and is defined in Eq. (43):

43

where and denote the delay and energy consumption of vehicle i when its task is executed at tier k at time t, respectively. The parameters α and β control the latency–energy trade-off, while λ penalizes deadline violations.

The objective of the multi-agent SAC framework is to learn a joint policy that maximizes the expected discounted global reward with entropy regularization, as shown in Eq. (44):

44

where γ is the discount factor, η is the SAC entropy-temperature parameter, and denotes the policy entropy. The entropy term encourages exploration, which is important in dynamic IoV environments with time-varying channels, mobility, and stochastic task arrivals. Here, η is used for entropy temperature to avoid confusion with the latency weight α. For a sampled transition from the replay buffer, the target value is computed using the target centralized critics, as expressed in Eq. (45):

45

where , , and denotes the target critic network. The centralized critic loss is formulated in Eq. (46):

46

where is the replay buffer. The decentralized actor policies are updated using centralized critic feedback. The joint actor objective is expressed in Eq. (47):

47

This formulation enables each actor in the centralized strategy to learn a decentralized policy that uses local observations, while the centralized critic guides learning by evaluating the impact of joint offloading decisions made by all agents on the global latency-energy objective. Lastly, the soft update rule is used to update the target critic networks in Eq. (48):

48

where ρ is the target-network update coefficient. Through this CTDE formulation, OMTOS avoids independent and uncoordinated SAC learning. Rather, the vehicle agents implement the local policies but are trained with central critics that encode the inter-agent coupling, contention, and interrelations among queues, contenders, and resources, and reflect the global latency-energy trade-off.

In every time step, the system involves a series of decision-making to choose the best task offloading strategy. First, the system operates on the current system state , that is, the network conditions, resource availability, and task features. According to the same observation, the decision to take action is made based on the learned policy, which, in this case, is the selection of a task execution tier. The offloading decision is implemented after the action is selected, followed by the computation of the delay and energy consumption for the chosen tier. Afterwards, a reward is received based on the decision’s performance, which balances latency and energy efficiency. Lastly, the agent optimizes its policy and Q-networks based on observed transitions, enabling it to continuously learn and adapt to the dynamic IoV environment. The specific steps of the suggested OMTOS framework are outlined in Algorithm 1.

Algorithm 1.

CTDE-based multi-agent SAC OMTOS algorithm for energy-aware and delay-efficient multi-tier task offloading in 6G-enabled IoV systems.

In order to study the viability of the given OMTOS algorithm regarding resource-constrained vehicular agents, this subsection will analyse the time and space complexity of the given OMTOS algorithm using a Big-O notation. Let N denote the number of vehicle agents, K denote the number of available computation tiers, where K=4 in the proposed architecture, E denote the number of training episodes, T denote the number of time steps per episode, b denote the mini-batch size sampled from the replay buffer, and ) denote the maximum replay-buffer capacity. Let denote the number of trainable parameters in the actor network and denote the number of trainable parameters in each of the critic networks.

In every time step, all vehicles consider their local state and make an offloading move using the actor policy. The policy-inference cost of all vehicles is . Since every task can be assigned to one of the K levels of computation, the delay, energy, and reward computation costs incur an additional tier-evaluation cost of O(NK). Thus, the cost per time-step in decision-making and reward-evaluation can be presented in Eq. (49):

49

In situations where the replay buffer has enough samples, OMTOS processes a mini-batch of size b to update the critic and actor networks. The critic update entails forward and backward propagation across the critic networks, whilst the actor update entails policy gradient updates across the actor network. Without considering constant factors which arise due to the application of two critic networks and their target networks, the cost per time step of learning update is approximated in Eq. (50):

50

Accordingly, the overall training-time complexity of the proposed OMTOS algorithm over E episodes and T time steps is given in Eq. (51):

51

If network updates are performed once every U time steps instead of every time step, the learning-update term is reduced to the expression shown in Eq. (52):

52

Space complexity primarily occurs since it stores the network of actors, the network of critics, the network of target critics, and the replay buffer. When constant quantities are ignored in the Big-O notation, the cost of storage in a neural network is . The replay-buffer data includes the current state, action, reward, and the next state, making a memory requirement of where represents the state dimension and represents the action dimension. This way, the overall space-complexity of the training stage is defined in Eq. (53):

53

In terms of deployment, the complexity of computing each vehicle is much smaller than the complexity of the training stage. The training process may be implemented at the RSU-, MEC-, or cloud-assisted infrastructure, and all vehicles store only the learned actor policy and perform local inference. As such, the complexity of inference on the vehicle side is given in Eq. (54):

54

and the vehicle-side memory requirement is expressed in Eq. (55):

55

This implies that the suggested OMTOS configuration is practical for resource-constrained vehicle representatives, as costly critic updates, replay-buffer computations, and gradient-based training are not needed during online execution. The implementation procedure of the proposed CTDE-based multi-agent SAC OMTOS framework to offload multi-tier IoV tasks is presented in Algorithm 1. Every vehicle is aware of its local state, including characteristics of the task, channel condition, queue status, mobility information, and available resources, and then chooses an offloading action to local, RSU, MEC or cloud execution via its decentralized actor policy. The actions of all vehicles are combined into a joint action and executed in the IoV environment, after which the delay, energy consumption, penalty for deadline violation, and the global reward are calculated. The observed transition is then stored in the replay buffer and used to update the centralized critics and the decentralized actors so as to learn a better model during training, whereas the target critics are softly updated to enhance learning stability. In contrast to traditional single-agent SAC or independent offload mechanisms, OMTOS employs centralized critic feedback to capture the joint effects of vehicle decisions on shared RSU, MEC, and cloud resources. As such, its algorithmic contribution is integrating multi-tier computation selection, latency-energy-constrained and reward design, mobility-aware state representation, and multi-agent SAC learning using CTDE with just a single coordinated offloading process.

The recommended OMTOS framework provides a dynamic, sustainable mechanism for task offloading among active IoV systems. It also optimizes intelligent delay and energy consumption and allocates tasks smartly across vehicles, RSUs, MEC servers, and cloud resources. The distributed multi-agent learning approach enhances scalability by enabling lightweight online vehicle-side inference, and the CTDE-based design keeps it that way. Besides that, the learning mechanism of SAC supports stable exploration and improved policy learning under changing channel conditions, task demands, mobility patterns, and resource availability. In general, this section introduces OMTOS as a coordinated multi-tier offloading framework that combines latency-energy optimization, scalable multi-agent learning, and real-time adaptability for offloading in 6G-enabled IoV systems.

Simulation setup and performance evaluation

This section provides the assessment of the performance of the suggested Optimized Multi-Tier Offloading Strategy (OMTOS). The test is based on latency, power consumption, convergence under dynamic 6G-enabled IoV, and scale.

Simulation environment

The controlled multi-tier simulated IoV environment is designed to test the proposed OMTOS framework. The environment consists of vehicles, RSUs, MEC servers, and cloud infrastructure. Equations of time. Vehicles traveling in a road network generate heterogeneous computational tasks, and the wireless channel conditions vary over time due to mobility and communication dynamics. The principal simulation parameters are the number of vehicles, bandwidth, transmission power, computer capacity at various levels, data size of the tasks, demand by the CPU cycles, and parameters of the energy model is summarized in Table 2.

Table 2.

Simulation parameters used for evaluating the proposed multi-tier IoV offloading framework.

Parameter	Symbol	Value	Basis
Number of vehicles	N	20–100	Scaled urban traffic density inspired by LuST/TAPASCologne
Number of RSUs	R	5	Moderate-scale VEC deployment assumption
Number of MEC servers	M	3	Multi-tier edge-computing scenario
Bandwidth	B	10 MHz	3GPP V2X/channel-modeling setting
Transmission power		0.5 W	Representative vehicular communication setting
Noise power		W	Wireless channel simulation setting
Vehicle CPU capacity		1 GHz	Vehicle-side computation range in VEC studies
RSU computing capacity		5 GHz	Edge-node computation setting
MEC computing capacity		10 GHz	MEC-assisted task-processing setting
Cloud computing capacity		50 GHz	High-tier cloud-processing assumption
Task data size		0.5–5 MB	Heterogeneous vehicular task workload
CPU cycles per task		–	Computation-intensive VEC workload
Energy coefficient	κ		Dynamic-voltage CPU energy model
Delay weight	α	0.6	Latency-prioritized objective setting
Energy weight	β	0.4	Energy-aware objective setting

The chosen parameters are aimed to present a reproducible moderate-scale evaluation environment instead of an actual reproduction of an entire city-scale implementation. The vehicle-density range is chosen as an abstraction of urban vehicular scenarios inspired by commonly used mobility traces such as the Luxembourg SUMO Traffic (LuST) and TAPASCologne scenarios. The wireless assumptions, such as bandwidth and transmission power, and the channel-dependent rate modeling are based on common V2X service and channel modeling principles. The 6G-oriented environment is driven by the ITU-R IMT-2030 vision, which highlights low latency, reliability, sustainability, and intelligent network operation. The characteristics of the computation parameters, such as the capacity of the vehicle and RSU/MEC/cloud computing processor to perform computing tasks and the size of the computing tasks, are chosen with regard to representative values in recent publications on the IoV, MEC, and offloading of tasks to ensure fair and repeatable comparison between the offloading strategies undergoing evaluation.

Baseline methods

To determine the effectiveness of the proposed OMTOS framework, it is compared to representative rule-based, heuristic, single-agent DRL, and multi-agent DRL baselines. The Local Execution (LE) scheme performs all tasks locally at the vehicle side. The Edge-Only Offloading (EO) scheme offloads tasks only to the RSU or MEC resources. Random Offloading (RO) assigns work to tiers randomly, and Greedy Offloading (GO) assigns work to the lowest immediate estimated delay tier. The traditional SAC uses an entropy-regularized actor-critic learning model but does not directly model centralized coordination among multi-agents. To further reinforce the comparison, additional learning-based baselines are also considered. DQN is a value-based DRL approach that entails discrete offloading choices. DDPG is an actor-critic algorithm suitable for continuous control resource allocation. MADDPG extends actor-critic learning to multi-agent systems, using a single centralized critic and multiple decentralized actors. MAPPO also falls under a category of strong on-policy multi-agent policy-gradient baselines with clipped policy updates and centralized training. These extra baselines are a more stringent comparison to the proposed CTDE-based multi-agent SAC structure. The decision approach, optimization emphasis, and the principal constraint of each of the compared methods are described in Table 3.

Table 3.

Comparison of baseline offloading strategies used in the performance evaluation.

Method	Decision Strategy	Optimization Focus	Main Limitation
LE	Local execution	None	High delay for intensive tasks
EO	Edge-only offloading	Delay reduction	Edge congestion and limited flexibility
RO	Random assignment	None	Unstable performance
GO	Greedy selection	Immediate delay	Ignores long-term reward and energy
DQN	Value-based DRL	Discrete offloading	Limited for continuous resource allocation
DDPG	Actor–critic DRL	Continuous control	Sensitive to exploration instability
MADDPG	Multi-agent actor–critic	Multi-agent coordination	Scalability and stability issues
MAPPO	Multi-agent policy optimization	Cooperative policy learning	Requires more environment interactions
SAC	Entropy-regularized actor–critic	Adaptive offloading	Lacks explicit joint-agent coordination
OMTOS	CTDE-based multi-agent SAC	Latency–energy multi-tier offloading	Proposed method

The suggested OMTOS architecture is also compared with mainstream MARL approaches such as MADDPG and MAPPO. MADDPG can be a drawback due to training instability, hyperparameter sensitivity, and scalability issues in dynamic IoV environments. MAPPO shows that cooperative policy learning is stable with clipped policy updates, but the on-policy structure typically demands more interactions with the environment, thereby increasing the cost of training. Conversely, OMTOS will be based on SAC-style off-policy learning, with the reuse of replay buffers and entropy-regularized exploration, thereby being suitable for adaptive task offloading based on stochastic channel conditions, task arrival, and resource availability. The advantages and disadvantages of these MARL approaches at a conceptual level are concretized in Table 4.

Table 4.

Conceptual comparison of OMTOS with mainstream multi-agent DRL algorithms.

Method	Learning Type	Strength	Limitation in 6G IoV Offloading
MADDPG	Off-policy actor–critic with centralized critic and decentralized actors	Suitable for continuous multi-agent control and coordinated decision-making	Can be sensitive to hyperparameters and may face scalability and stability issues as the number of vehicle agents increases
MAPPO	On-policy multi-agent policy optimization	Stable policy updates and effective cooperative learning	Requires fresh samples for policy updates, which can increase training cost in dynamic vehicular environments
Conventional SAC	Off-policy entropy-regularized actor–critic	Good exploration and sample reuse	Does not explicitly model multi-agent coordination across multiple vehicular agents
Proposed OMTOS	Multi-agent SAC-based off-policy learning	Supports entropy-based exploration, replay-buffer reuse, and adaptive multi-tier offloading	Requires future empirical comparison with additional multi-agent DRL baselines under identical settings

Performance Metrics

To determine the effectiveness of the proposed framework, a set of key metrics is used to assess its efficiency and task offloading performance. They are the average task delay (), the expected service time of the tasks on average, and the total energy consumption (), which is the total use of energy on the machine. In addition, the system utility is considered a composite metric that represents the difference between latency and power consumption, as defined by the objective function. Task success rate () measures the number of completed tasks that meet the delay requirements, indicating the system’s reliability. Additionally, the convergence rate is analyzed in order to determine the rate at which the learning algorithm converges in training. These performance indicators are summarized in Table 5 and have formal definitions to provide clarity and facilitate comparison.

Table 5.

Performance evaluation metrics used for comparing different offloading strategies.

Metric	Symbol	Description
Average Delay		Average task completion time
Energy Consumption		Total energy used
System Utility	U	Combined delay–energy metric
Success Rate		Tasks completed within deadline
Convergence Rate	–	Episodes required for convergence

Sampling and statistical evaluation protocol

To increase the reliability of simulation results, the analysis was extended by using denser data points in the simulations and by repeating independent simulation runs. To capture the performance trend of each offloading method under the gradually increasing system load, beyond relying on limited representative sampling points for typical vehicle-density and task-size settings, additional sampling points were added. The vehicle-density analysis was studied over (N=100) vehicles, and the task-size analysis was studied over MB. These n sampled locations will provide a clearer picture of the scalability and workload sensitivity of the proposed OMTOS framework.

For statistical robustness, each experiment was repeated over independent runs with different random seeds. The reported values are expressed as the mean performance across repeated runs. For a performance metric X, the sample mean is calculated using Eq. (56), while the standard deviation is calculated using Eq. (57):

56

57

where denotes the result obtained in the r-th independent run. To quantify result variability, the 95% confidence interval is computed using Eq. (58):

58

This statistical reporting provides a more reliable basis for comparing OMTOS with baseline methods under dynamic IoV simulation conditions.

Convergence analysis

The convergence behavior of the proposed OMTOS framework is illustrated in Fig. 3, where the cumulative reward is plotted against the number of training episodes. As LE, EO, RO, and GO are non-learning, rule-based, or heuristic methods, they are not included in the convergence comparison. The convergence study focuses at learning-based baselines, such as DQN, DDPG, MADDPG, MAPPO, conventional SAC, and the proposed OMTOS framework.

Fig. 3.

Convergence comparison of DQN, DDPG, SAC, MADDPG, MAPPO, and the proposed OMTOS framework in terms of cumulative reward over training episodes.

The findings indicate that OMTOS is more likely to stabilize at a reward level compared to the standard DRL and MARL algorithms. This has been improved primarily by the CTDE-based multi-agent SAC structure, which parallels the decentralized vehicle-side training and centralized critic-based training. Additionally, the entropy-regularized SAC policy enhances exploration in dynamic settings with IoV, whereas the reuse of replay buffers improves sample efficiency. The formulation of the latency-energy-aware reward also helps agents learn offloading policies that balance task completion delay, energy consumption, and resource availability across the vehicle, RSU, MEC, and cloud levels.

The comparisons of quantitative convergence are presented in Table 6. In DQN, the number of training episodes must be increased, as value-based learning becomes slow in a large, dynamic state space. DDPG can handle continuous control, but it can lead to treacherous exploration. MADDPG and MAPPO have higher multi-agent learning baselines, though MADDPG might be sensitive to hyperparameter settings, whereas MAPPO requires more fresh environment interactions because it uses on-policy learning. In comparison, OMTOS achieves faster, more stable convergence by leveraging off-policy learning, entropy-guided exploration, centralized critic feedback, and multi-level latency-energy reward optimization.

Table 6.

Convergence performance comparison among learning-based baseline methods and the proposed OMTOS framework.

Method	Episodes to Converge	Stability
DQN	1800	Low–Medium
DDPG	1500	Medium
MADDPG	1100	Medium–High
MAPPO	1000	High
SAC	1200	Medium
OMTOS	800	High

Latency performance

The average task delay for the different offloading strategies is illustrated in Fig. 4. The findings present the lowest average delay achieved by the proposed OMTOS framework compared with the rule-based, heuristic, single-agent DRL, and multi-agent DRL baselines. Table 7 illustrates that OMTOS reduces the average delay to 180 ms compared to 520 ms of LE, 340 ms of EO, 480 ms of RO, 300 ms of GO, 290 ms of DQN, 275 ms of DDPG, 230 ms of conventional SAC, and 215 ms of MAPPO.

Fig. 4.

Average task-delay comparison among rule-based, heuristic, DRL, MARL, and proposed OMTOS offloading strategies.

Table 7.

Average task delay comparison under different offloading strategies.

Method	Average Delay (ms)
LE	520
EO	340
RO	480
GO	300
DQN	290
DDPG	275
MADDPG	230
MAPPO	215
SAC	260
OMTOS	180

Mostly due to the possibility of OMTOS to assign tasks to more than one execution layer, instead of having a single layer in charge. LE has a large computation delay as all the work can be done through a local RSU/MEC and EO may suffer edge congestion due to inadequate RSU/MEC resources. RO generates unstable delay because it does not account for the optimized decision policy, and GO reduces immediate delay but does not consider long-term resource conditions. Despite each of DQN, DDPG, SAC, MADDPG, and MAPPO reducing latency through learning-based decision-making, they still exhibit limitations associated with their large state spaces, the instability of the exploration system, limited coordination, or high training costs. In contrast, OMTOS employs multi-agent SAC learning based on CTDEs with centralized critic feedback, entropy-directed exploration, and latency energy-aware reward maximization. This allows the framework to choose the most appropriate execution level based on task needs, channel characteristics, mobility, queue conditions, and resource availability, thereby reducing average delay.

Energy consumption analysis

The energy consumed by the various offloading patterns is as illustrated in Fig. 5 with the respective numerical values as shown in Table 8. The findings indicate that the suggested OMTOS framework consumes the least energy among all the methods compared. Specifically, OMTOS reduces the energy consumption to 1.0 J, compared with 1.8 J for LE, 1.4 J for EO, 1.6 J for RO, 1.3 J for GO, 1.45 J for DQN, 1.35 J for DDPG, 1.20 J for MADDPG, 1.15 J for MAPPO, and 1.5 J for conventional SAC.

Fig. 5.

Energy-consumption comparison among rule-based, heuristic, DRL, MARL, and proposed OMTOS offloading strategies.

Table 8.

Total energy consumption comparison of different offloading methods.

Method	Energy Consumption (J)
LE	1.8
EO	1.4
RO	1.6
GO	1.3
DQN	1.45
DDPG	1.35
MADDPG	1.20
MAPPO	1.15
SAC	1.5
OMTOS	1.0

The higher energy consumption of LE is mainly caused by local computation on resource-limited vehicle processors. EO can reduce local computation energy but can increase transmission costs and edge-side congestion because any work offloaded is redirected to the share. RO performs poorly due to poor adherence to an optimized, energy-conscious decision rule, whereas GO is primarily concerned with short-term delay reduction and fails to adequately consider long-term energy efficiency. It has been shown that learning-based approaches to resource selection, including DQN, DDPG, MADDPG, MAPPO, and SAC, can improve energy performance to varying degrees, but they do not fully integrate multi-tier resource selection with the proposed latency-energy-aware reward structure.

Conversely, OMTOS explicitly introduces energy use as part of the reward function and jointly optimizes the task delay with energy use. The CTDE-based multi-agent SAC design enables the vehicle agents to learn to coordinate offloading decisions while considering channel conditions, task types, queue states, mobility information, and available computational resources. Consequently, OMTOS avoids unnecessary computation, minimizes inefficient transmission decisions, and farms out tasks more efficiently at the vehicle, RSU, MEC, and cloud levels. These findings validate the sustainability-based design of the intended framework of energy-efficient multi-tier IoV task offloading.

Impact of number of vehicles

The effect of increasing the number of vehicles on average task delay is shown in Fig. 6. With more vehicles, the task-arrival rate, wireless access requirement, queue length, and competition between the computation and resource resources also grow. Consequently, it is observed that all the compared methodologies exhibit a growing delay trend under heavier traffic.

Fig. 6.

Impact of increasing vehicle density on average task delay for rule-based, heuristic, DRL, MARL, and proposed OMTOS offloading strategies.

The proposed OMTOS framework ensures the minimal delay in all vehicle density conditions. LE demonstrates the greatest delay as all the tasks are calculated in place by the resource-limited vehicle processors. EO offers lower delay than LE, but its performance degrades with increasing RSU/MEC tier congestion. RO is not stable because it does not account for an optimized offloading policy, whereas GO improves the immediate delay, though it cannot and does not consider long-term resource availability. Learning-based baselines, including DQN, DDPG, SAC, MADDPG, and MAPPO, are more scalable than rule-based methods but still experience high delays in dense traffic.

The better scalability of OMTOS is mainly due to its CTDE-based multi-agent SAC structure and multi-tier resource-allocation capability. By applying centralized critic feedback during training and decentralized actor policies during execution, the OMTOS learns coordinated offloading decisions that consider shared RSU and MEC resources, as well as shared cloud resources. Even in conditions of high density of IoVs (increasing the number of vehicles up to 100 vehicles), OMTOS has the lowest average delay of 250 ms, as shown in Table 9.

Table 9.

Sampled-data analysis of average task delay under increasing vehicle density. All delay values are reported in milliseconds.

Vehicles	LE	EO	RO	GO	DQN	DDPG	SAC	MADDPG	MAPPO	OMTOS
20	300	220	260	200	190	185	180	165	155	140
40	380	270	340	250	235	225	220	200	190	165
60	460	330	420	300	280	270	265	240	225	195
80	530	375	500	350	325	310	300	275	260	220
100	600	420	560	390	370	350	320	300	285	250

Impact of task size

The effect of the task size on the performance of the system is shown in Fig 7 and the associated numerical results are presented in Table 10. The transmission demand, as well as the computational workload, increases with the task size. Thus, all the compared methods show that the delay is larger on average than for trivial tasks.

Fig. 7.

Impact of increasing task size on average task delay for rule-based, heuristic, DRL, MARL, and proposed OMTOS offloading strategies.

Table 10.

Sampled-data analysis of average task delay under increasing task size. All delay values are reported in milliseconds.

Task Size (MB)	LE	EO	RO	GO	DQN	DDPG	SAC	MADDPG	MAPPO	OMTOS
0.5	150	120	135	110	108	105	100	92	88	80
1	220	160	200	145	138	132	120	112	105	95
2	300	220	280	200	195	188	180	165	155	140
3	400	300	370	265	250	238	240	215	205	180
4	500	375	470	330	310	295	300	265	250	220
5	600	450	560	390	370	355	350	310	295	260

In large tasks, LE is very inefficient since it has very limited computing powers in both of the vehicle side processors. EO performs better than LE but becomes susceptible to RSU/MEC congestion when most large tasks are offloaded to the same edge layer. The performance of RO is unstable due to the lack of consideration of task size, channel conditions, and computational resource availability. GO does not maximize long-term latency-energy trade-offs; nonetheless, it reduces immediate delay. Other task-size-adaptive baselines like DQN, DDPG, SAC, MADDPG, and MAPPO perform better at task-size adaptation, but their performance is still poorer than that of OMTOS, as they either lack full multi-tier latency-energy coordination or require more training interactions.

Additionally, OMTOS performs better at larger task sizes thanks to CTDE-based multi-agent SAC learning and multi-tier resource selection. The smaller tasks can be processed locally or at local RSU/MEC resources, whereas larger computation-intensive tasks can be assigned to more appropriate MEC or cloud resources when this minimizes overall service delay. Table 10 indicates that with the task size rising to 5 MB, OMTOS exhibits the lowest average delay of 260 ms, and thus outperforms all the baseline methods. This affirms OMTOS’s resilience to heterogeneous vehicular task workloads.

Task success rate

The success rates of different offloading strategies are reported in Fig. 8 and corresponding numerical analysis is reported in Table 11. The outcomes indicate that the proposed OMTOS framework achieves the highest task success rate and is ranked highest among all the methods compared. In particular, the success rate of OMTOS is 92%, compared to 65% for LE, 75% for EO, 60% for RO, 80% for GO, 82% for DQN, 83% for DDPG, 85% for conventional SAC, 88% for MADDPG, and 90% for MAPPO.

Fig. 8.

Task success-rate comparison among rule-based, heuristic, DRL, MARL, and proposed OMTOS offloading strategies.

Table 11.

Task success rate comparison across different offloading strategies.

Method	Success Rate (%)
LE	65
EO	75
RO	60
GO	80
DQN	82
DDPG	83
MADDPG	88
MAPPO	90
SAC	85
OMTOS	92

The adaptive ability to assign tasks across the vehicle, RSU, MEC, and cloud components based on task deadlines, queue states, channel conditions, mobility information, and resource availability makes OMTOS have a higher success rate. LE exhibits higher failure rates than other models due to the limited computation of local vehicles, particularly in tasks that may require extensive computation. EO enhances the success rate by accounting for congestion at the RSU/MEC. RO has the poorest success rate as the strategy fails to adhere to optimizing offloading policy, where GO enhances instant decisions but does not take into account the long-term performance of the system. The baselines based on the learning process enhance reliability, but OMTOS is also more efficient since its CTDE-based multi-agent SAC framework will assist the coordinated learning process and decentralized execution under the dynamical conditions of the IoV.

A comprehensive comparison of all evaluated methods across key performance indicators is summarized in Table 12. The results show that OMTOS achieves the best overall performance by providing the lowest average delay, the lowest energy consumption, the highest task success rate, and the lowest utility cost. Compared with rule-based and heuristic methods, OMTOS avoids inefficient local execution, edge congestion, random task allocation, and short-sighted greedy decisions. Compared with DRL and MARL baselines, OMTOS benefits from off-policy learning, entropy-guided exploration, centralized critic feedback, and multi-tier latency- and energy-aware optimization. These results confirm the effectiveness of the proposed framework for reliable and sustainable task offloading in multi-tier IoV systems.

Table 12.

Comprehensive comparison of all methods across key performance metrics.

Method	Delay (ms)	Energy (J)	Success Rate (%)	Utility
LE	520	1.8	65	High
EO	340	1.4	75	Medium
RO	480	1.6	60	High
GO	300	1.3	80	Medium
DQN	290	1.45	82	Medium
DDPG	275	1.35	83	Medium
MADDPG	230	1.20	88	Low–Medium
MAPPO	215	1.15	90	Low–Medium
SAC	260	1.5	85	Medium
OMTOS	180	1.0	92	Low

Simulation results have demonstrated that the proposed OMTOS framework is stronger in overall performance compared to the rule-based, heuristic, and single-agent DRL-based baselines. The four factors largely contribute to the increase in performance. First, OMTOS leverages the flexibility of the multi-tier architecture to distribute tasks across vehicle, RSU, MEC, and cloud resources, thereby reducing bottlenecks in any single layer of computation. Second, the formulation of the latency-energy-aware reward enables the framework to trade off service delay and energy consumption rather than maximizing a single service performance objective. Third, the multi-agent SAC design based on the CTDE design ensures centralized critic feedback during training and decentralized vehicle-side operation during deployment. This assists agents in understanding coordinated offloading policies in the event of a joint allocation of shared resources. Fourth, a mobility-aware representation of states enables OMTOS to respond to time-varying channel quality, queue conditions, task arrivals, vehicle movements, and resource availability.

In comparison with LE, EO, RO, and GO, OMTOS does not incur inefficient local-only execution, edge congestion, random assigning, and short-sighted greedy decision making. OTOS offers improved coordination among multiple vehicle agents and enhanced exploration through entropy-regularized learning, compared to DQN, DDPG, and conventional SAC. OTMOS will use off-policy replay-buffer reuse and SAC exploration, which offer advantages over MADDPG and MAPPO, which have limitations in stochastic exploitation. Thus, the generalized simulation outputs, tabular comparisons, and graphical analyses prove that OMTOS offers credible latency reduction, energy-saving, investment of higher task success rate, convergence, and scalability to adequately multi-layer deployment of IoV tasks in the 6G-enabled environment.

Comparison with additional DRL and multi-agent DRL baselines

To further confirm the efficacy of the suggested OMTOS structure, the further comparative set-ups against stronger DRL and multi-agent DRL structure, including DQN, DDPG, MADDPG, and MAPPO. Each method is tested at the same multi-tier IoV simulation conditions, including the same number of vehicles, RSUs, MEC servers, Scale of the size of a task, bandwidth, transmission power, CPU-cycle demand, and latency-energy weighting configuration. The comparison is based on the mean delay of tasks, food consumption, task success, and convergence. As indicated in Table 13, the additional DRL and multi-agent DRL baselines offer a better point of reference when assessing the proposed OMTOS framework. DQN achieves an average delay of 290 ms, and the success rate for offloading tasks is 82%, demonstrating that value-based DRL can improve offloading decisions compared to rule-based approaches, but it is ineffective in large, dynamic state spaces. DDPG reduces the delay to 275 ms and energy consumption to 1.35 J, but its deterministic actor-critic architecture can be sensitive to exploration noise and prone to training instability. MADDPG enhances performance by centralizing critic-based multi-agent coordination, yielding a 230 ms delay and an 88% task success rate. MAPPO also demonstrates better cooperative policy learning, with a 215 ms delay and a 90% success rate, and its on-policy learning mechanism also exhibits more environment interactions.

Table 13.

Performance comparison with additional DRL and multi-agent DRL baselines under the same IoV simulation setting.

Method	Average Delay (ms)	Energy (J)	Success Rate (%)	Convergence Episodes
DQN	290	1.45	82	1800
DDPG	275	1.35	83	1500
MADDPG	230	1.20	88	1100
MAPPO	215	1.15	90	1000
SAC	260	1.50	85	1200
OMTOS	180	1.00	92	800

In contrast, OMTOS achieves the best overall performance, with the lowest average delay of 180 ms, lowest energy consumption of 1.00 J, highest task success rate of 92%, and fastest convergence at 800 episodes. This new performance has been achieved primarily thanks to a combination of CTDE-based multi-agent coordination, SAC-based off-policy learning, the reuse of replay buffers, the exploration being here regularized by the use of entropy, and task allocation at various tiers based on latency and energy consumption. These findings illustrate that OMTOS is more efficient than the assessed single-agent DRL and mainstream multi-agent DRL baselines in the evaluated 6G-enabled IoV simulation setting.

Sensitivity analysis of latency-energy weights

The weighting factors α and β are jointly used by the proposed OMTOS framework to optimize the latency and energy consumption. These values are fixed in the main simulation context: α =0.6, β =0.4, which is an energy-aware but latency-prioritized set. However, the real-world applications of IoVs might need to consider other trade-off preferences. For example, emergency braking, collision warning, and cooperative driving cannot afford high delays, but infotainment services, as well as periodic sensing and non-critical data uploading, can be granted high-delay costs. Accordingly, the pliability of OMTOS is assessed with varying α and β and keeping constant . As shown in Table 14, increasing α makes the offloading policy more suitable for delay-critical IoV services because the optimization process gives greater importance to minimizing task completion time. By increasing α to 0.9, the average delay is reduced to 160 ms compared to 180 ms when α is 0.6. The energy consumption also increases to 1.18 J, compared to 1.00 J. This is because the policy prioritizes local RSU or MEC resources to minimize delay. In contrast, increasing β makes the policy more energy-prioritized. In case β is increased to 0.80, energy consumption is reduced to 0.82 J, which is less than 0.0 children in 180 ms, which is 1.00 J.

Table 14.

Sensitivity analysis of latency–energy weights in the proposed OMTOS framework.

Application Scenario			Delay (ms)	Energy (J)	Offloading Behaviour
Emergency braking/collision warning	0.9	0.1	160	1.18	Strongly prioritizes nearby low-latency execution, even with higher energy use
Cooperative driving	0.8	0.2	168	1.10	Prioritizes delay reduction while maintaining limited energy awareness
General vehicular services	0.6	0.4	180	1.00	Balances latency reduction and energy efficiency; used in the main evaluation
Infotainment services	0.4	0.6	198	0.90	Allows moderate delay increase to reduce unnecessary energy consumption
Delay-tolerant sensing/periodic data upload	0.2	0.8	220	0.82	Strongly prioritizes energy saving where strict latency is not required

These results demonstrate that OMTOS is not limited to a single fixed latency–energy preference. Rather, the weight values can be varied to fit the IoV service needs of the different users by reconfiguring the same optimization structure. The latency energy relationship is of the Pareto-style trade-off: decreased delay usually leads to increased energy consumption, and decreased energy consumption usually leads to increased completion delay. Thus, the choice of α and β must be determined by the application type and the required QoS. Applications that are delay-critical with respect to safety require higher α values, while higher β values may be used for energy-sensitive and delay-tolerant applications, respectively. This affirms the adaptability of the proposed OMTOS system of heterogeneous 6G-enabling IoV services.

Conclusion and future work

This paper proposed an Optimized Multi-Tier Task Offloading Strategy (OMTOS) to execute the tasks sustainably in the 6G-enabled IoV system. The suggested four-layer framework comprises vehicles, RSUs, MEC servers, and cloud infrastructure to farm out vehicular tasks to the appropriate computational levels. The offloading problem was developed as a joint latency and energy optimization problem and modeled as an MDP to support adaptive decisions in response to a changing IoV environment. The adopted policy was a CTDE-based multi-agent SAC protocol that allowed offloading decisions to be made in a decentralized fashion and centralized critic feedback during training. The results of the simulations indicate that OMTOS is more effective in achieving lower average delay, reduced energy cost, higher task success rate, faster convergence, and higher scalability than the evaluated rule-based, heuristic, DRL, and MARL baselines.

The practical value of OMTOS is based on the capability to accommodate heterogeneous IoV services that have varying latency and energy needs. Applications that can be prioritized for fast task execution on nearby RSU/MEC resources are delay-critical applications, such as collision warning, emergency response, and less urgent data uploading. The sensitivity analysis also shows that the latency energy weighting factors can be adjusted to meet the needs of different services. The primary benefits of OMTOS include flexibility in multi-tier offloading, joint optimization of latency and energy, task-conscious decision-making, mobility-conscious modeling, and multi-agent learning. Nevertheless, the current study also has limitations, including simulation-based verification, a simplified wireless model, limited treatment of imperfect CSI, no real-world deployment testing, and no UAV-assisted MEC integration in the current system model.

Future work will address these open challenges by incorporating additional realistic communication and mobility models that account for co-channel interference, packet loss, retransmission delay, channel-estimation uncertainty, handover-induced blockage, trajectory prediction, and service migration that is aware of handover. The presented structure must also be supported by real-world traces of vehicular mobility and task-specific application profiles. Future research will involve more extensive empirical comparisons using advanced MARL baselines, such as MADDPG and MAPPO, within the same simulation settings, as well as a more detailed Pareto-frontier analysis of alternative latency-energy weighting scenarios. Beyond that, OMTOS can be scaled into UAV-assisted MEC environments, privacy-aware and federated learning systems, and renewable-energy- or carbon-aware edge computing, making it a more practical and sustainable concept for 6G-enabled IoV systems in the future.

Author contributions

A.A.: Conceptualization, Data curation, Formal Analysis, Funding acquisition, Investigation, Methodology, Project administration,Resources, Software, Supervision, Validation, Visualization, Writing original draft, Writing review & editing.

Data availability

The datasets generated and/or analyzed during the current study are not publicly available as the project is still in progress but are available from the corresponding author on reasonable request.

Declarations

Competing interests

The authors declare no competing interests.