Oranits: Open RAN ITS Mission Assignment Framework

• Oranits is a unified optimization framework for mission assignment and task offloading in intelligent transportation systems, integrating autonomous vehicles, MEC, and cloud resources.
• It employs a hybrid approach combining the Chaotic Gaussian-based Global ARO metaheuristic with a Multi-agent Double Deep Q-Network for dynamic scheduling.
• The framework jointly addresses mission dependencies, offloading costs, and scheduling constraints within the Open RAN architecture to maximize deadline-feasible mission completions.

Searching arXiv for the cited Oranits paper and closely related Open RAN/ITS task-offloading work to ground the article. Oranits is a framework for mission assignment and task offloading in Open RAN-based intelligent transportation systems (ITS), introduced to coordinate autonomous vehicles, mobile edge computing (MEC), and cloud resources under mission deadlines, mission interdependence, and offloading budget constraints (Nguyen et al., 25 Jul 2025). It addresses a joint optimization problem in which missions must be assigned to vehicles, ordered on each vehicle, and supported by edge/cloud computation through an Open RAN architecture, while explicitly accounting for predecessor–successor relations among missions and the communication/computation costs of offloading (Nguyen et al., 25 Jul 2025). The framework combines a one-slot metaheuristic solver, Chaotic Gaussian-based Global ARO (CGG-ARO), with a dynamic deep reinforcement learning method, Multi-agent Double Deep Q-Network (MA-DDQN), to improve mission completion and system benefit in dynamic ITS environments (Nguyen et al., 25 Jul 2025).

1. Concept and operational scope

Oranits is defined as a unified optimization framework and system model for mission assignment, mission scheduling, and task offloading in an Open RAN + MEC ITS environment (Nguyen et al., 25 Jul 2025). Its target setting is one in which autonomous vehicles execute transportation or logistics missions while generating onboard computational workloads that may be offloaded to MEC servers or a cloud server through Open RAN infrastructure (Nguyen et al., 25 Jul 2025). The model explicitly incorporates mission interdependence, task offloading costs, and vehicle cooperation, rather than treating missions or offloading decisions in isolation (Nguyen et al., 25 Jul 2025).

The framework is built around an Open RAN deployment with one cloud server $\mathcal{S}^c$, multiple MEC servers $\mathcal{S}^m$, distributed units at MEC servers, centralized units in the cloud, and radio units attached to distributed units as vehicle access points (Nguyen et al., 25 Jul 2025). Decision intelligence is split between the Near-RT RIC, which supports near-real-time control and xApps, and the Non-RT RIC, which supports offline AI/ML training and policy guidance (Nguyen et al., 25 Jul 2025). This architectural placement is significant because it links mission scheduling and task offloading to native Open RAN control loops rather than treating them as an external optimization service (Nguyen et al., 25 Jul 2025).

The practical objective is to maximize the number of missions completed before deadlines while respecting dependency and budget constraints (Nguyen et al., 25 Jul 2025). A key implication is that Oranits is not only a routing or offloading model; it is a coupled scheduling-and-computation framework in which transportation execution and compute placement are evaluated jointly (Nguyen et al., 25 Jul 2025).

2. System model and mathematical formulation

Over a scheduling period $\tau$, the available vehicles form the set $\mathcal{K}(\tau)$ with $|\mathcal{K}(\tau)|=K$, and a subset of size $K^*$ is selected for a mission batch based on proximity to missions via an xApp in the Near-RT RIC (Nguyen et al., 25 Jul 2025). The server set is $\mathcal{S}=\mathcal{S}^m\cup\mathcal{S}^c$, where $\mathcal{S}^m$ denotes MEC servers and $\mathcal{S}^c$ the cloud server, with the chosen offloading server written $S_o\in\mathcal{S}$ (Nguyen et al., 25 Jul 2025).

The mission set arriving over $\mathcal{S}^m$0 is partitioned into $\mathcal{S}^m$1 subsets of size $\mathcal{S}^m$2, organized as a mission matrix $\mathcal{S}^m$3 in the paper’s notation, with optimization performed row-by-row (Nguyen et al., 25 Jul 2025). Each mission is modeled as

$\mathcal{S}^m$4

where $\mathcal{S}^m$5 is the route, $\mathcal{S}^m$6 the deadline, $\mathcal{S}^m$7 the offloading budget, and $\mathcal{S}^m$8 the predecessor and successor mission sets (Nguyen et al., 25 Jul 2025). This dependency representation is central: if $\mathcal{S}^m$9, the mission may start immediately; if $\tau$0, it does not constrain subsequent missions (Nguyen et al., 25 Jul 2025).

Each mission $\tau$1 has a task set $\tau$2, and each task $\tau$3 is characterized by $\tau$4, where $\tau$5 is the input data size in bits and $\tau$6 the required CPU cycles (Nguyen et al., 25 Jul 2025). The return data size is assumed negligible (Nguyen et al., 25 Jul 2025). Traffic conditions are abstracted into five road states—free flow, stable flow, slow flow, congested flow, and severe congestion—with route selection obtained via Dijkstra’s algorithm (Nguyen et al., 25 Jul 2025).

For each mission subset row $\tau$7, the decision variable is

$\tau$8

where $\tau$9 is the assigned vehicle and $\mathcal{K}(\tau)$0 the execution order on that vehicle (Nguyen et al., 25 Jul 2025). The formulation therefore simultaneously determines vehicle assignment and intra-vehicle mission ordering (Nguyen et al., 25 Jul 2025).

3. Constraints, latency model, and objective function

The formulation imposes four assignment-structure constraints. First, each mission row is handled by exactly $\mathcal{K}(\tau)$1 vehicles: $\mathcal{K}(\tau)$2 Second, each mission is assigned to one vehicle: $\mathcal{K}(\tau)$3 Third, each mission has at most one scheduling order on a vehicle: $\mathcal{K}(\tau)$4 Fourth, distinct missions assigned to the same vehicle must have distinct orders: $\mathcal{K}(\tau)$5 These are complemented by dependency-order constraints: $\mathcal{K}(\tau)$6

$\mathcal{K}(\tau)$7

All of these relations are explicit in the Oranits formulation (Nguyen et al., 25 Jul 2025).

Offloading follows a greedy policy: a vehicle queries reachable MEC servers, estimates latency to each, selects the MEC server with the lowest latency, then compares that choice with the cloud and picks the lower-latency option (Nguyen et al., 25 Jul 2025). The wireless uplink rate for task $\mathcal{K}(\tau)$8 of mission $\mathcal{K}(\tau)$9, sent by vehicle $|\mathcal{K}(\tau)|=K$0 to server $|\mathcal{K}(\tau)|=K$1, is

$|\mathcal{K}(\tau)|=K$2

with $|\mathcal{K}(\tau)|=K$3 the channel bandwidth, $|\mathcal{K}(\tau)|=K$4 the vehicle transmit power, and $|\mathcal{K}(\tau)|=K$5 the noise spectral density (Nguyen et al., 25 Jul 2025). Cloud access adds fiber delay

$|\mathcal{K}(\tau)|=K$6

while communication delay is

$|\mathcal{K}(\tau)|=K$7

Total communication delay is $|\mathcal{K}(\tau)|=K$8, and total computation delay is

$|\mathcal{K}(\tau)|=K$9

where $K^*$0 is the server computational capacity (Nguyen et al., 25 Jul 2025). Travel delay is

$K^*$1

and the isolated-mission delay is

$K^*$2

The paper assumes these movement, communication, and computation delays are non-overlapping (Nguyen et al., 25 Jul 2025).

Mission completion time incorporates queueing and dependency chains: $K^*$3 where $K^*$4 contains earlier missions on the same vehicle, $K^*$5 predecessor missions assigned to other vehicles, and $K^*$6 missions preceding those predecessors on their own vehicles (Nguyen et al., 25 Jul 2025). This is one of the framework’s distinguishing features, because it propagates mission-order effects across vehicles rather than only within a single execution queue (Nguyen et al., 25 Jul 2025).

The offloading cost of mission $K^*$7 is

$K^*$8

with budget feasibility expressed as

$K^*$9

The optimization problem $\mathcal{S}=\mathcal{S}^m\cup\mathcal{S}^c$0 is

$\mathcal{S}=\mathcal{S}^m\cup\mathcal{S}^c$1

subject to all assignment, ordering, dependency, completion-time, and budget constraints (Nguyen et al., 25 Jul 2025). The paper states that $\mathcal{S}=\mathcal{S}^m\cup\mathcal{S}^c$2 is NP-hard via reduction to deadline scheduling on a single machine (Nguyen et al., 25 Jul 2025).

4. Optimization methods: CGG-ARO and MA-DDQN

Oranits proposes a twofold optimization strategy: a metaheuristic baseline for one-slot optimization and a DRL method for repeated dynamic decision making (Nguyen et al., 25 Jul 2025). The first component is CGG-ARO, explicitly named Chaotic Gaussian-based Global ARO, an enhanced Artificial Rabbits Optimization method used as a one-slot solver for $\mathcal{S}=\mathcal{S}^m\cup\mathcal{S}^c$3 (Nguyen et al., 25 Jul 2025). The second is MA-DDQN, a Multi-agent Double Deep Q-Network designed for dynamic mission arrivals and near-real-time scheduling adaptation (Nguyen et al., 25 Jul 2025).

In CGG-ARO, a candidate solution for population member $\mathcal{S}=\mathcal{S}^m\cup\mathcal{S}^c$4 at generation $\mathcal{S}=\mathcal{S}^m\cup\mathcal{S}^c$5 is encoded as

$\mathcal{S}=\mathcal{S}^m\cup\mathcal{S}^c$6

where $\mathcal{S}=\mathcal{S}^m\cup\mathcal{S}^c$7 is a mission index permutation and $\mathcal{S}=\mathcal{S}^m\cup\mathcal{S}^c$8 a vehicle index vector, both of length $\mathcal{S}=\mathcal{S}^m\cup\mathcal{S}^c$9 (Nguyen et al., 25 Jul 2025). Each vehicle index appears exactly $\mathcal{S}^m$0 times, and the chromosome is decoded into the assignment-order representation $\mathcal{S}^m$1 (Nguyen et al., 25 Jul 2025). To improve initialization diversity, CGG-ARO uses a Piecewise Chaotic Map: $\mathcal{S}^m$2 It then uses Gaussian exploration,

$\mathcal{S}^m$3

and an exploitation update

$\mathcal{S}^m$4

where $\mathcal{S}^m$5 is an opposition-based term and $\mathcal{S}^m$6 a global-best attraction term (Nguyen et al., 25 Jul 2025). An enhanced random hiding stage is also introduced (Nguyen et al., 25 Jul 2025). The stopping criterion is $\mathcal{S}^m$7, and the worst-case time complexity is

$\mathcal{S}^m$8

with the fitness-evaluation cost $\mathcal{S}^m$9 dominating in practice (Nguyen et al., 25 Jul 2025).

MA-DDQN is formulated with $\mathcal{S}^c$0 agents corresponding to $\mathcal{S}^c$1 vehicles (Nguyen et al., 25 Jul 2025). It is deployed with offline training in the Non-RT RIC and online inference as an xApp in the Near-RT RIC, with the trained policy delivered via the A1 interface (Nguyen et al., 25 Jul 2025). The observation of agent $\mathcal{S}^c$2 at decision step $\mathcal{S}^c$3 is

$\mathcal{S}^c$4

combining road state, vehicle state, mission-assignment memory, and mission information (Nguyen et al., 25 Jul 2025). The action space is the mission index set $\mathcal{S}^c$5, with $\mathcal{S}^c$6-greedy selection

$\mathcal{S}^c$7

The paper gives Bellman-style expressions

$\mathcal{S}^c$8

$\mathcal{S}^c$9

and states that the DDQN implementation uses a target network, shared replay buffer, and off-policy mini-batch learning (Nguyen et al., 25 Jul 2025).

A central element of MA-DDQN is its enhanced composite reward: $S_o\in\mathcal{S}$0 where the shared reward is

$S_o\in\mathcal{S}$1

and the dependency reward is

$S_o\in\mathcal{S}$2

The paper explicitly contrasts this reward with simple mission-success rewards, arguing that the latter are too myopic because they ignore dependency relief, shared system benefit, and budget-awareness (Nguyen et al., 25 Jul 2025).

5. Evaluation, empirical behavior, and trade-offs

The Oranits evaluation is conducted on a $S_o\in\mathcal{S}$3 area around VinUniversity campus, Hanoi, Vietnam, centered at $S_o\in\mathcal{S}$4, with $S_o\in\mathcal{S}$5 vehicles, up to $S_o\in\mathcal{S}$6 missions, a scheduling window of $S_o\in\mathcal{S}$7 minutes, $S_o\in\mathcal{S}$8 MEC servers, and $S_o\in\mathcal{S}$9 cloud server (Nguyen et al., 25 Jul 2025). Network parameters include $\mathcal{S}^m$00 MHz cellular bandwidth, $\mathcal{S}^m$01 mW transmission power, $\mathcal{S}^m$02 RU antennas, path loss exponent $\mathcal{S}^m$03, $\mathcal{S}^m$04 uplink channels, noise PSD $\mathcal{S}^m$05 dBm/Hz, and fiber rate $\mathcal{S}^m$06 Gbps (Nguyen et al., 25 Jul 2025). The baselines are APO, SHADE, L-SHADE, EO, ARO, CGG-ARO, and MA-DDQN (Nguyen et al., 25 Jul 2025).

For the metaheuristic experiments, the settings are $\mathcal{S}^m$07 iterations, population $\mathcal{S}^m$08, and $\mathcal{S}^m$09 seeds (Nguyen et al., 25 Jul 2025). CGG-ARO achieves fitness $\mathcal{S}^m$10, completed missions $\mathcal{S}^m$11, and total benefits $\mathcal{S}^m$12 (Nguyen et al., 25 Jul 2025). The abstract reports that CGG-ARO improves the number of completed missions and the overall benefit by approximately $\mathcal{S}^m$13 and $\mathcal{S}^m$14, respectively, over baseline methods (Nguyen et al., 25 Jul 2025). The paper also notes better median fitness and benefit in boxplots and faster convergence than compared metaheuristics (Nguyen et al., 25 Jul 2025).

For DRL, the modified-reward MA-DDQN converges to total system benefit around $\mathcal{S}^m$15 after about $\mathcal{S}^m$16 training epochs, whereas the unmodified reward converges near $\mathcal{S}^m$17 (Nguyen et al., 25 Jul 2025). Across 15 mission sets, MA-DDQN attains the highest mission count in $\mathcal{S}^m$18 sets and the highest fitness in $\mathcal{S}^m$19 sets (Nguyen et al., 25 Jul 2025). The abstract states that MA-DDQN improves mission completions by $\mathcal{S}^m$20 and overall benefit by $\mathcal{S}^m$21 over baseline approaches (Nguyen et al., 25 Jul 2025).

The empirical record also shows that MA-DDQN is not uniformly dominant. In mission set 9, CGG-ARO obtains $\mathcal{S}^m$22 completed missions and fitness $\mathcal{S}^m$23, while MA-DDQN reaches $\mathcal{S}^m$24 completed missions and fitness $\mathcal{S}^m$25 (Nguyen et al., 25 Jul 2025). By contrast, in mission set 4, CGG-ARO yields $\mathcal{S}^m$26 missions and fitness $\mathcal{S}^m$27, whereas MA-DDQN yields $\mathcal{S}^m$28 missions and fitness $\mathcal{S}^m$29 (Nguyen et al., 25 Jul 2025). This supports the paper’s stated trade-off: CGG-ARO can be stronger in highly structured one-slot scenarios, while MA-DDQN is preferable for fast online adaptation in dynamic settings (Nguyen et al., 25 Jul 2025).

A major operational distinction is runtime. CGG-ARO is reported to require about $\mathcal{S}^m$30 iterations, with $\mathcal{S}^m$31–$\mathcal{S}^m$32 seconds per iteration, for a total of roughly $\mathcal{S}^m$33 minutes, making it unsuitable for real-time use (Nguyen et al., 25 Jul 2025). MA-DDQN, by contrast, incurs training overhead but provides near-real-time inference after deployment (Nguyen et al., 25 Jul 2025). This division of labor between optimization quality and deployment latency is one of the framework’s central practical themes (Nguyen et al., 25 Jul 2025).

6. Significance, limitations, and prospective extensions

Oranits is positioned as a response to three deficiencies in earlier ITS and offloading work: neglect of mission interdependence, neglect of offloading cost, and failure to jointly integrate mission scheduling, Open RAN support, MEC/cloud task placement, and multi-agent coordination (Nguyen et al., 25 Jul 2025). The paper’s claimed contributions are the Oranits framework itself, the optimization problem formulation, the CGG-ARO one-slot solver, the MA-DDQN dynamic extension, and the comparative evaluation against state-of-the-art metaheuristics (Nguyen et al., 25 Jul 2025).

Its significance lies in tying mission-level vehicular coordination to the Open RAN control stack. The Non-RT RIC trains and updates the policy, while the Near-RT RIC executes near-real-time scheduling decisions as an xApp (Nguyen et al., 25 Jul 2025). This suggests a broader interpretation of Open RAN intelligence as supporting application-layer vehicular orchestration in addition to radio optimization. A plausible implication is that Oranits can be read as an instance of AI-native Open RAN control for cyber-physical ITS workloads, although the paper’s evidence is confined to its simulation setting (Nguyen et al., 25 Jul 2025).

The framework also has explicit simplifying assumptions. Missions are optimized in fixed-size batches of $\mathcal{S}^m$34; a fixed number $\mathcal{S}^m$35 of nearby vehicles is preselected; offloading uses a greedy latency-based policy rather than fully joint compute-placement optimization; communication, computation, and movement delays are assumed non-overlapping; feedback data size is negligible; and traffic state is fixed during each $\mathcal{S}^m$36 (Nguyen et al., 25 Jul 2025). The paper shows empirical convergence for CGG-ARO and MA-DDQN but provides no formal convergence guarantees (Nguyen et al., 25 Jul 2025).

The reported future directions are federated learning for privacy-preserving adaptation across distributed nodes and the combination of DRL with graph neural networks for more complex interdependent mission scenarios (Nguyen et al., 25 Jul 2025). These are natural extensions because the current formulation already depends heavily on relational structure among missions, vehicles, and servers. This suggests that the next stage of the line of work may strengthen relational inductive bias and distributed training support without changing the central Oranits objective: maximizing deadline-feasible mission completion under dependency and budget constraints in an Open RAN-based ITS (Nguyen et al., 25 Jul 2025).