PaperlayerView exact revision on arXiv

Source-linked AI summary

Place, Slice and Schedule: Hierarchical O-RAN Control of a Tethered mmWave UAV-gNB

Alireza Mohammadhosseini, Fatemeh Afghah

arXiv:2608.23824v1cs.NIcs.AI

TL;DR

The paper addresses the missing coordination between slow aerial topology control and fast per-user scheduling in O-RAN-enabled tethered UAV networks. It proposes a hierarchical rApp/xApp controller with a permutation-equivariant D-SAC scheduler and ray-traced mmWave channels, improving both eMBB and URLLC service metrics over baselines.

  • Problem

    Existing O-RAN-enabled UAV studies typically optimize trajectory, deployment, association, or resource allocation in isolation rather than coordinating slow aerial control with fast per-user scheduling.

  • Method

    A Non-RT rApp jointly controls tethered UAV placement and eMBB/URLLC slice PRB budgets, while a Near-RT permutation-equivariant DeepSets SAC xApp allocates per-user PRBs using ray-traced channel context.

  • Results

    Up to 17% higher eMBB SLA satisfaction and 42% higher URLLC on-time delivery are achieved over classical and learned schedulers, while the learned rApp raises URLLC on-time delivery by up to 20% over baselines.

  • Takeaways & Limitations

    Jointly designing the two RIC control loops around per-UE objectives improves service across both eMBB SLA satisfaction and URLLC on-time delivery.

Abstract

from arXiv · show

Unmanned aerial vehicle (UAV)-mounted 5G New Radio base stations (gNBs) can augment terrestrial networks with an on-demand, repositionable Frequency Range 2 (FR2) capacity layer. This flexibility, however, couples the physical network topology with radio-resource management: UAV movement reshapes blockage, channel quality, and the set of effectively served users, while traffic demand, queues, and service requirements evolve at a much faster timescale. Existing Open Radio Access Network (O-RAN)-enabled UAV studies optimize trajectory, deployment, association, or resource allocation, but typically in isolation, without coordinating slow aerial control with fast per-user scheduling. We instead exploit O-RAN disaggregation, Key Performance Indicator (KPI) monitoring, and multi-timescale RAN Intelligent Controller (RIC) control to address this coupling: a Non-Real-Time RIC rApp uses aggregated KPIs and radio-environment context to jointly control tethered UAV placement and the enhanced Mobile Broadband (eMBB)/Ultra-Reliable Low-Latency Communication (URLLC) slice budget, while a Near-Real-Time RIC xApp allocates per-user resources within that budget. We realize this xApp as a permutation-equivariant DeepSets Soft Actor-Critic (D-SAC) scheduler that treats the users as an unordered set, trained in a Sionna RT ray traced channel. The resulting hierarchical controller improves eMBB SLA satisfaction by up to 17% and URLLC on-time delivery by up to 42% over classical and learned schedulers; the learned rApp further raises URLLC on-time delivery by up to 20% over baselines.

I. INTRODUCTION

The paper addresses coupled slow UAV placement and slice control alongside fast per-user scheduling in a tethered mmWave O-RAN system. It proposes a hierarchical controller with a permutation-equivariant D-SAC scheduler and evaluates it using ray-traced FR2 channels.

  • Motivation: Slow UAV placement and slice-budget decisions interact with fast per-user allocation because movement changes channel quality, blockage, coverage, and service outcomes.The two control loops therefore cannot be tuned independently.
  • Proposed controller: The Non-RT rApp jointly selects tethered UAV 3D placement and the eMBB/URLLC slice PRB split, while the Near-RT xApp allocates PRBs per user.The rApp policy is handed to the xApp as an A1 policy.
  • Evaluation: The controller is evaluated with Sionna RT ray-traced FR2 channels that capture line-of-sight and blockage geometry.Channel gains for tethered UAV candidates are precomputed rather than generated solely from a stochastic channel model.
  • Scheduler: D-SAC uses shared per-UE processing and permutation-invariant pooling to support variable UE populations without dependence on UE ordering or explicit graph construction.The scheduler is designed to preserve the natural symmetry of unordered user sets.
  • Results: The hierarchical controller improves eMBB SLA satisfaction and URLLC on-time delivery over classical and learned baselines.The evaluation also reports that D-SAC retains its performance advantage as the UE population increases without retraining.

II. RELATED WORK

Prior work studies aerial placement, association, trajectories, offloading, slicing, and resource allocation, but generally does not jointly coordinate controllable aerial topology with fast per-user scheduling. The paper combines multi-timescale O-RAN control with a set-based scheduler and ray-traced channel modeling.

  • UAV control: Earlier UAV studies commonly optimize placement for coverage, association, throughput, or another aggregate objective while fixing or abstracting the fast scheduler.This leaves the coupling between UAV movement and the feasible per-user service region insufficiently addressed.
  • O-RAN control: O-RAN UAV studies jointly consider trajectories, deployment, association, offloading, or resource allocation, but their resource decisions remain tied primarily to those objectives.The additional feedback loop created by a controllable aerial gNB is not addressed in the cited prior work.
  • Set-based scheduling: Set-based scheduling represents active UEs as an unordered set using shared processing and permutation-invariant pooling, enabling operation across UE counts independently of ordering.This is presented as a simpler alternative to explicit graph construction when users mainly compete for a shared resource budget.
  • System design: The proposed system uses a slow Non-RT rApp for UAV placement and slice-level PRB budgets and a fast Near-RT xApp for conditional per-UE allocation.The rApp generates the A1 policy, while the xApp uses E2SM-RC and E2SM-KPM interfaces in the described O-RAN system.
  • Channel model: The channel model uses tether-feasible UAV candidates and Sionna RT-derived location-dependent gains, with slot-scale fading and PRB-based downlink rates.The candidate geometry is tethered to a fixed ground anchor and the downlink is modeled as a single UAV-gNB cell.

B. Service Model and QoS Metrics

The service model evaluates heterogeneous eMBB minimum-rate guarantees and deadline-constrained URLLC delivery under shared slice PRB budgets. It combines aggregate QoS shaping with per-UE SLA and deadline metrics.

  • Service model: The rApp selects eMBB and URLLC slice budgets, while the xApp allocates those budgets among individual UEs at each radio slot.The mean eMBB throughput is computed over the radio intervals in each control window.
  • eMBB metrics: eMBB SLA violation occurs when a UE’s throughput falls below its own minimum-rate guarantee Ru,min.The aggregate eMBB QoS score instead uses a network-level reference rate Rtarget for reward shaping.
  • URLLC model: URLLC is modeled as packet-based, deadline-constrained GBR traffic with per-UE packet delay budgets in the tens of milliseconds.Packet arrivals follow a Poisson model, and queued packets may be retransmitted while deadlines and retransmission budgets permit.
  • URLLC metrics: URLLC on-time delivery is reported as a scheduling metric under an intentionally overloaded regime rather than as a reliability figure.Deadline misses are counted when packets are not delivered within Du seconds of arrival.

IV. PROBLEM FORMULATION

The controller is formulated as a hierarchical MDP with separate slow rApp and fast xApp decisions. Each layer observes information at the granularity appropriate to its control timescale.

  • Hierarchical formulation: The hierarchical MDP splits control into a slow rApp decision and a fast xApp scheduling decision.The rApp uses aggregate KPIs, while the xApp uses finer per-UE radio-control information.

A. Hierarchical MDP

The hierarchical MDP separates slow UAV placement and slice-budget control from fast per-UE PRB allocation. The rApp uses aggregated window-level state, while the xApp operates on per-UE slot-level features within the selected budgets.

  • rApp state and action: The rApp selects a tethered UAV movement command and eMBB/URLLC slice split, determining the per-slice PRB budgets for the next control window.Movement is restricted to one tether-feasible grid hop per control step.
  • rApp state and action: The rApp state aggregates placement, movement, slice, UE spatial, radio-quality, QoS-pressure, composition, and neighboring-position service features over the previous 1 s window.This matches KPI-level information available to a Non-RT RIC rather than per-slot MAC state.
  • xApp state and action: The xApp observes a per-UE feature matrix combining channel, queue, throughput, deadline/SLA state, and rApp slice-budget context, then emits one continuous action per UE.Those actions are mapped to feasible PRB allocations within the rApp-selected slice budgets.
  • Policy factorization: The hierarchical policy makes the xApp conditional on rApp intent through slice-budget context and the resulting channel and queue state.The continuous PRB-share relaxation represents average shares over the radio-control interval; integer PRB rounding is outside the model.

B. Objective and Reward

The objective combines eMBB and URLLC QoS while penalizing SLA violations and deadline misses. Because the UAV is tethered, mobility is constrained by feasible movement rather than an energy term.

  • Reward design: The window-level QoS score is a weighted sum of eMBB QoS and URLLC QoS.The resulting reward also includes penalties for eMBB and URLLC violations.
  • Mobility constraints: The tethered UAV objective includes no movement-energy term because power is supplied through the cable.Mobility is instead constrained through the feasible tether graph.
  • Reward design: The per-step reward subtracts weighted eMBB SLA-violation and URLLC deadline-miss penalties from the QoS score.Training uses fixed positive penalty weights as a Lagrangian relaxation of the constrained objective.

V. PROPOSED SOLUTION

The proposed solution trains the hierarchical controller in stages to avoid the computational cost of end-to-end optimization. The xApp is first exposed to varied high-level policies, followed by rApp training and scheduler adaptation.

  • Staged training: End-to-end optimization is computationally expensive because each rApp action induces many xApp decisions.The staged procedure preserves the hierarchy during training.
  • Staged training: The xApp is first trained as a conditional scheduler under randomized UAV candidates and slice budgets.This trains fast-timescale scheduling across varied high-level policies.
  • Staged training: The xApp is then frozen while the rApp is trained on the induced slow-timescale MDP, after which the xApp is fine-tuned with the learned rApp frozen.Fine-tuning adapts the scheduler to the state distribution induced by the rApp.

A. Per-UE PRB Scheduling

The xApp uses a DeepSets SAC policy to allocate PRBs among unordered UEs while respecting rApp-selected slice budgets. Shared processing and invariant pooling provide permutation-equivariant per-UE priorities without explicit graph construction.

  • Scheduling formulation: The conditional xApp allocates per-UE PRBs every slot within fixed UAV and slice-budget choices from the rApp.Its reward combines eMBB QoS and URLLC queue QoS while penalizing SLA violations and deadline misses.
  • Permutation-equivariant policy: DeepSets SAC avoids dependence on UE ordering by using shared per-UE processing and permutation-invariant global pooling.This supports variable-cardinality UE populations without explicit graph construction or message passing.
  • Actor architecture: Mean pooling captures average traffic and channel conditions, while max pooling captures extreme urgency or backlog.A shared stochastic SAC head combines each UE embedding with the global context to produce one action per UE.
  • Critic architecture: The critic pools per-UE feature/action representations and predicts one scalar Q-value for the whole allocation.Thus, the actor is equivariant over UE outputs while the critic is invariant over the unordered set.

C. rApp Training over the Frozen xApp

The rApp is trained as a slow-timescale MDP over aggregate O-RAN state while the frozen xApp, mobility, arrivals, and channel variation define its environment. Experiments use a tethered UAV with ray-traced candidate positions and discrete eMBB/URLLC slice splits.

  • C. rApp Training over the Frozen xApp: The frozen per-UE scheduler becomes part of the slow-timescale MDP transition kernel used to train the rApp.The rApp applies Double DQN over joint tether-movement and slice-budget actions.
  • C. rApp Training over the Frozen xApp: The deployment area is 200 m × 200 m, with a 100 m tether and 21 candidate UAV positions spanning 64.3–100 m.Candidates combine five elevation and five azimuth levels; channel-gain maps are precomputed at 28 GHz using Sionna RT.
  • C. rApp Training over the Frozen xApp: The rApp selects among five eMBB/URLLC slice splits ranging from (0.9, 0.1) to (0.1, 0.9).UEs move at 1 m/s under a uniform random walk.
  • C. rApp Training over the Frozen xApp: Heterogeneous eMBB loads, minimum-rate SLAs, URLLC arrival rates, and deadlines create unequal per-UE service pressure.Some eMBB targets exceed cell capacity, while ray-traced blockage can reduce channel gain to the noise floor.

VII. RESULTS

The results compare high-level rApp policies and low-level schedulers using per-UE service metrics. Learned placement is the dominant rApp contribution, while D-SAC achieves the strongest reported eMBB and URLLC performance among scheduler baselines.

  • VII. RESULTS: Learned placement reaches 0.440 eMBB satisfaction and 0.687 URLLC on-time delivery, exceeding split-only values of 0.391 and 0.594.The full rApp further raises eMBB satisfaction to 0.463 while URLLC on-time delivery is 0.682.
  • VII. RESULTS: The full learned rApp surpasses the strong greedy method on both eMBB SLA satisfaction and URLLC on-time delivery.The comparison varies only the high-level policy while keeping the same D-SAC scheduler.

2) Scheduler Comparison:

D-SAC provides the strongest combined service performance and remains effective as the UE population grows, while other schedulers expose throughput–fairness or scalability trade-offs.

  • Scheduler performance: 0.463 eMBB SLA satisfaction and 0.682 URLLC on-time delivery are the highest reported values for D-SAC among the compared schedulers.D-SAC exceeds PF, Max-Rate, Equal, and SAC-MLP on the reported per-UE service metrics.
  • Service tradeoff: D-SAC trades aggregate-rate fairness for per-UE SLA satisfaction under heterogeneous minimum-rate contracts.Max-Rate achieves the highest throughput but the worst URLLC on-time delivery and lowest fairness, whereas Equal is fairest but throughput-limited.
  • Scalability: Across 30, 40, and 50 UEs, D-SAC retains the highest eMBB SLA satisfaction and URLLC on-time delivery without retraining.The reported eMBB values are 0.46, 0.44, and 0.39; URLLC values are 0.68, 0.58, and 0.55.
  • Overall comparison: D-SAC improves eMBB SLA satisfaction by up to 17% and URLLC on-time delivery by up to 42% over classical and learned schedulers.The conclusion identifies D-SAC as the only scheduler performing highly on both service axes.
  • Controller-level result: The learned rApp further improves URLLC on-time delivery by up to 20% over random placement and surpasses a greedy-coverage heuristic.The scheduler also retains its lead from 30 to 50 users without retraining.
Loading 2608.23824v1…