PaperlayerView exact revision on arXiv

Source-linked AI summary

WONDER: A Radio World Model-based Negotiation Framework for Multi-Agent UAV Coverage Optimization

Jiahao Huang, Rongpeng Li, Zhifeng Zhao, Guoru Ding, Honggang Zhang

arXiv:2608.16955v1cs.MAcs.LG

TL;DR

Post-disaster UAV coverage restoration must coordinate swarm-level trajectories despite hidden radio consequences and sparse inter-UAV communication. WONDER combines a JEPA-based radio world model with multi-round negotiation, achieving the highest balanced score among evaluated methods on 11 testing scenes and 100% gateway connectivity.

  • Problem

    Deployable multi-UAV coverage restoration lacks effective coordination under hidden radio consequences and sparse inter-UAV communication.

  • Method

    WONDER predicts candidate trajectories’ incremental radio consequences with a JEPA-based world model and coordinates them through multi-round negotiation.

  • Results

    0.870 balanced score was the highest among evaluated methods on 11 RadioDynamics testing scenes, with 100% gateway connectivity.

  • Takeaways & Limitations

    The results support WONDER as a deployable framework for multi-UAV coverage restoration under hidden radio effects and sparse communication.

Abstract

from arXiv · show

Post-disaster damage to terrestrial infrastructure can disrupt wireless coverage,while Uncrewed Aerial Vehicle (UAV) swarms provide a promising solution for rapid restoration.However, due to the limitations in local geometry observations hidden radio impact,and inter-UAV communication,there exists a significant gap between locally visible movement choices and swarm-level coverage outcomes.To combat this gap,we propose a raido World-model-based Optimized Negotiation framework for Distributed UAV covERage (WONDER).Particularly, to tackle the unavailability of the future radio field from onboard observations, WONDER uses a Joint-Embedding Predictive Architecture (JEPA)-based radio world model to learn and predict the incremental radio effect of each candidate trajectory from deployment-available information.Multi-round negotiation in WONDER then coordinates ranked proposals by committing one trajectory at a time and re-evaluating the remaining proposals under the updated context. Our theoretical analyses further validate the effectiveness of such a world model-based framework. WONDER also adopts a Proximal Policy Optimization (PPO)-style Actor and alternates between updating the world model and the actor. Furthermore,we build RadioDynamics,a comprehensive simulation environment that integrates UAV mobility,radio propagation, inter-UAV communication modeling,and digital-twin geometry with ray-traced fields in $62$ metropolitan scenes.Experiments on $11$ testing scenes in RadioDynamics show that WONDER achieves the highest balanced score among seven evaluated methods,reaching $0.870$ with a $0.162$ coverage advantage over STACCA, while maintaining $100\%$ connectivity between UAVs.

I. INTRODUCTION … A. System Model

The paper frames post-disaster UAV coverage restoration as a decentralized problem involving hidden radio effects, partial observations, limited communication, and connectivity constraints. It introduces WONDER, a JEPA-based radio world model with multi-round proposal negotiation, and evaluates it in the RadioDynamics environment.

  • I. INTRODUCTION: Post-disaster infrastructure damage creates an urgent need for rapid wireless coverage restoration, for which UAV swarms provide flexible, resilient connectivity.UAV swarms establish adaptive connectivity graphs to serve end users.
  • B. Motivation and Contributions: Coverage expansion must preserve inter-UAV connectivity despite unavailable radio feedback, local observations, and limited nearby communication.The system must relate visible geometry and candidate movements to service effects that are hidden at decision time.
  • A. Related Works: Prior work spans UAV emergency networking, MARL coordination under partial observations, and radio-aware deployment using maps, geometry, and coverage prediction.These research lines motivate integrating cooperative control with propagation-dependent service quality.
  • B. Motivation and Contributions: WONDER predicts each candidate trajectory’s incremental radio impact with a JEPA-based radio world model and coordinates ranked proposals through sequential multi-round negotiation.Remaining proposals are reevaluated after each negotiated commitment under the evolving context.
  • B. Motivation and Contributions: WONDER alternates updates between its radio world model and PPO actor, while analysis addresses the counterfactual composition gap and derives a recursive lower bound for sequential plans.The bound becomes attainable with proposal coverage and bounded election regret.
  • B. Motivation and Contributions: RadioDynamics integrates UAV mobility, radio propagation, inter-UAV communication, digital-twin geometry, and ray-traced fields across 62 metropolitan radio scenes.The scenes represent 700 m × 700 m central urban districts and use realistic three-dimensional building geometry.
  • C. Paper Organization: The paper organizes its development from system modeling and problem formulation through negotiation theory, WONDER methodology, proof, and experimental evaluation.The system model precedes the diagnosis of synchronized counterfactual deployment and the proposed negotiation framework.
  • A. System Model: The system model defines a finite-horizon UAV-swarm restoration problem in which UAV positions induce an A2G service field and an A2A communication graph over a digital-twin service region.At each deployment step, decentralized actions use local observations and coordination messages, while geometry-based ray tracing determines aggregate A2G power gain.

1) Communication Model:

The communication model defines ground-service fields from serving-UAV signal power and characterizes inter-UAV communication through a geometry- and propagation-dependent A2A graph. Backhaul connectivity is obtained through direct-access roots and multi-hop paths, with synchronization requiring full gateway connectivity.

  • A2G Ground Service: The serving UAV is selected by maximum received power, and its field defines signal strength Rt(x), link quality γt(x), and rate rt(x).
  • A2A Communication: The A2A graph (Nl, El) specifies one-hop message exchange between each UAV and its neighbors.Neighbor relations depend on the deployment layout, pairwise distance, carrier frequency, effective antenna height, and LoS/NLoS propagation.
  • A2A Communication: An A2A edge is included only when received power exceeds the threshold ρlink in both directions.This bidirectional condition supports reliable coordination between UAVs.
  • Backhaul Connectivity: UAVs with direct backhaul access form the root set Nroot, while other UAVs obtain core-network access through multi-hop A2A paths to a root.A UAV is gateway-connected when it has a multi-hop A2A path to at least one root.
  • Backhaul Connectivity: At synchronization states, the hard gateway-connectivity requirement is C(St; e) = 1.

2) Decision Model: · B. Problem Formulation

The decision model restricts each UAV to deployment-available local observations and discrete movements, while coordinating proposals through periodic neighbor communication. After heuristic flocking warm-up, the decentralized policy seeks to maximize a weighted connectivity-preserved coverage objective over seven evaluation metrics.

  • 2) Decision Model:: Each UAV observes a 160-dimensional one-hop representation containing shared geometry, its position, communication topology, neighbor positions, and neighboring geometry.The input is formed by concatenating five 32-dimensional group representations.
  • 2) Decision Model:: The policy cannot access the global radio field or states of non-neighbor UAVs.This limitation motivates decision-making from deployment-available local information.
  • 2) Decision Model:: Every H steps, synchronization triggers R communication rounds in which UAVs exchange neighbor messages containing their top-Kc trajectory proposals.Each proposal includes a trajectory, its latent representation, and an evaluation score.
  • B. Problem Formulation: During the Nwarm-step warm-up prefix, the swarm follows a heuristic flocking policy that disperses UAVs while maintaining inter-UAV connectivity.The action space includes four cardinal movements and hovering, with separation, crowding, and connectivity-degradation terms.
  • B. Problem Formulation: The heuristic warm-up rapidly disperses UAVs but leaves substantial optimization potential because it lacks radio-aware gain assessment.The limitation is explicitly associated with the resulting warm-up trajectory.
  • B. Problem Formulation: After heuristic initialization, decentralized policy π runs for Nlearn synchronization rounds and induces a seven-dimensional evaluation-metric vector µ(Sπ; e).The reported state Sπ is obtained after the complete restoration process.
  • B. Problem Formulation: The objective J optimizes connectivity-preserved coverage using evaluation weights w and seeks a policy π that maximizes the resulting objective.The spatial averaging operator is defined over the service area Ωs.

III. EMPIRICAL ANALYSIS · IV. METHODOLOGY · A. Overview

The empirical analysis identifies counterfactual composition errors in parallel UAV planning and motivates sequential, multi-round negotiation. WONDER implements this design by filtering, ranking, sharing, and electing trajectories to assemble joint plans.

  • III. EMPIRICAL ANALYSIS: WONDER addresses the composition gap by committing one trajectory per negotiation round and recomputing remaining proposals under the updated context.This sequential construction uses the same counterfactual evaluation as parallel planning while updating the reference plan after each commitment.
  • III. EMPIRICAL ANALYSIS: 82.1% of 3,726 active UAV pairs have positive counterfactual composition gaps, with a mean gap of 1.00% and median gap of 0.73%.A positive GCF(N; Tbase) indicates that individually improved trajectories overestimate the gain realized in parallel execution.
  • III. EMPIRICAL ANALYSIS: The sequential-plan advantage over the parallel plan is positive for every r ≥ 1, increasing from 0.047 at r = 1 to 0.261 at r = 5.All 95% confidence intervals remain positive, and the advantage grows with the negotiation round.
  • A. Overview: At each synchronization step, WONDER constructs a joint plan through multiple trajectory-election rounds and selects actions only after finalizing the plan.Each round commits one elected trajectory before the next round proceeds.
  • A. Overview: Each UAV filters endpoints reachable within H Manhattan-distance steps and constructs non-overlapping feasible trajectories τi,k ∈ Tfeas.Candidate trajectories are reduced before negotiation using deployment-available information Ii,t.
  • A. Overview: The JEPA radio world model ranks trajectories and retains the top Kc local candidates for subsequent negotiation.The model produces the retained set Tloc_i,r for each UAV.
  • A. Overview: Neighbor messages are reduced within subgraphs, relayed across subgraphs within at most Kdep hops, and combined into a global pool of Kc trajectories.The resulting global trajectory pool is provided to the PPO-based election actor Aϕ.
  • A. Overview: After R negotiation rounds, WONDER assembles the joint plan from committed trajectories and local proposals of UAVs that remain uncommitted.The elected trajectory becomes the committed trajectory and is appended to the committed-plan set for the next round.

B. Radio World Model

WONDER uses a radio world model to estimate each feasible trajectory’s decision-relevant radio impact from deployment-visible information. The model conditions predictions on evolving negotiation context and trains deployment representations against rollout-measured metric changes.

  • Motivation: The radio world model estimates each feasible trajectory’s decision-relevant radio impact when future radio fields and non-neighbor states are unavailable at deployment.It addresses the infeasibility of directly evaluating trajectories using only deployment information I_i,t.
  • Architecture: Wθ combines a deployment encoder, metric-change prediction head, and rollout-supervised target encoder.The deployment-time architecture is illustrated in Fig. 5 alongside the broader inference and training pipelines.
  • Contextual prediction: Across negotiation rounds, Wθ predicts each candidate trajectory’s incremental impact relative to trajectories already committed under the evolving context.The resulting trajectory score υ_i,k is computed from predicted metric change, with µ combining coverage area, Jain fairness, and connectivity.
  • Training objective: JEPA training supervises both target and deployment paths with rollout-measured metric changes and aligns deployment latents with stopped target latents.A coherent loss additionally encourages nearby trajectory endpoints to have consistent target latents.

C. Training Process · V. THEORETICAL ANALYSIS

WONDER separates radio-world-model representation learning from election learning, then alternates their refinement with refreshed proposals and rollouts. Theoretical results show that sequential negotiation increasingly improves its lower bound over a parallel counterfactual plan and approaches that advantage under proposal-effectiveness and bounded-regret assumptions.

  • C. Training Process: WONDER pretrains the JEPA radio world model Wθ and trajectory scorer Pθ on offline rollouts, freezes Wθ, and trains the election module Aϕ.The election rewards combine balanced-objective improvement with a gateway-connectivity penalty.
  • C. Training Process: Alternating refinement updates the proposal pool through Wθ, discards stale PPO rollouts, and trains Aϕ on fresh rollouts generated by the updated world model.This coupling repeatedly re-encodes and scores candidates under the evolving context.
  • V. THEORETICAL ANALYSIS: Sequential commitment yields a larger lower bound on J than the STACCA-style parallel counterfactual plan as negotiation rounds increase.The analysis compares Tseq with T∥ and explains how the learned policy approaches the bound in the practical negotiation range.
  • V. THEORETICAL ANALYSIS: When GCF_r ≥ 0, the sequential gain over T∥ is nondecreasing from round r to round r + 1; when GCF_r > 0, it increases strictly.This is the recursive advantage stated by Theorem 1 under Assumption 1.
  • V. THEORETICAL ANALYSIS: Inference forms feasible, non-overlapping trajectory proposals, reduces and relays messages across communication subgraphs, constructs Tglobal, and commits trajectories over multiple rounds.Algorithm 1 terminates when no active UAVs remain or negotiation ends.
  • V. THEORETICAL ANALYSIS: Iterating the recursion shows that sequential negotiation advantage over the parallel plan is reduced by cumulative selection regret Σ_{t=0}^{r−1}ϵ_t.The result applies when recursive composition gaps are nonnegative and proposal coverage is empirically verified.

VI. EXPERIMENT SETTINGS AND RESULTS · A. Experimental Setup

The experiments use RadioDynamics, built from 62 metropolitan scenes with ray-traced radio fields, and evaluate methods across separate training, validation, and test splits. Baselines differ in their available scene priors, local observations, and communication interfaces, while deployable methods cannot access ground-truth RSRP.

  • A. Experimental Setup: 62 metropolitan scenes, including Hong Kong, New York, and Tokyo, are used to construct the RadioDynamics environment.The digital-twin pipeline converts city data into scene assets for simulation.
  • A. Experimental Setup: OpenStreetMap geometry is converted into mesh-based city models, assigned empirical materials, and used with Sionna RT to precompute dense ground-level RSRP fields.Each scene stores transmitter settings together with its resulting RSRP fields.
  • A. Experimental Setup: The 62 scenes are split into 42 training scenes, 9 validation scenes, and 11 test scenes, with 4 entries per scene.Methods train with seed 0 and are evaluated with seeds {0, . . . , 5}.
  • A. Experimental Setup: 15 warm-up steps of Flocking spread every test swarm before the evaluated algorithm runs for 30 learning steps from the same initial state.The scene-aware A2A graph is updated from UAV positions, digital-twin line of sight, and link-budget information.
  • A. Experimental Setup: Separate radios support A2G service broadcast, A2A swarm relay, and ground backhaul, with each UAV carrying a 3.5-GHz, 100-MHz, 25-dBm-EIRP service-radio payload.The A2G payload is represented by a Canopy 5GID1 n78 small-cell radio envelope.
  • A. Experimental Setup: MAPPO, STACCA, and Flocking use shared static scene priors and deployment-visible local observations without multi-hop communication, whereas MAGI and WONDER use one-hop communication, topology, and selected trajectories.Ground-truth Reference Signal Received Power (RSRP) is unavailable to deployable methods.

B. Experiemental Results … VII. CONCLUSION

WONDER improves closed-loop UAV coverage through steady, connected trajectory adaptation, low communication overhead, and coordinated multi-round negotiation. Across RadioDynamics evaluations and ablations, its JEPA world model, election mechanism, alternating optimization, and negotiation design support the strongest reported performance.

  • 1) Performance Superiority:: WONDER steadily improves the score while preserving relay links, ending 0.413 above Flocking and 0.162 above STACCA in the representative rollout.Flocking disconnects at t = 35, whereas STACCA fluctuates due to counterfactual mismatch.
  • 2) Feasibility Study:: 270.336 kbit is WONDER’s total proposal-aggregation payload per synchronization round, including 147.456 kbit added over STACCA.The payload comparison covers local-vector exchange, proposal aggregation, and raw full-map transfer under the same topology.
  • 2) Feasibility Study:: 0.0676% of nominal round capacity is used by WONDER, keeping communication load within the bandwidth of the listed devices.A 1-s deployment step provides 400 Mbit over the 5-step synchronization round using an 80 Mbps A2A reference.
  • 3) Ablation Analysis:: 99.62% recall is achieved by JEPA at Kc = 3, outperforming DR by 4.92% and supporting reliable retention of subsequent globally best actions.JEPA and DR both converge during local proposal training.
  • 3) Ablation Analysis:: Proposal-level negotiation prioritizes actions with higher global gain than fixed-order greedy selection, avoiding the local optimum induced by Sequential Exhausted.The comparison concerns sequentially visiting UAVs and committing each visited UAV’s locally best action.
  • 3) Ablation Analysis:: 0.870 is the full method’s highest balanced score, 7.67% above the strongest ablation, while it is the only variant reaching 100% connectivity.The strongest ablation is the variant without alternating updates.
  • 3) Ablation Analysis:: Removing JEPA, learned election, alternating optimization, or multi-round negotiation reduces balanced score to 0.659, 0.708, 0.808, and 0.784, respectively.The ablations replace or remove the corresponding WONDER components under the shared evaluation setting.
  • VII. CONCLUSION: WONDER combines a JEPA-based radio world model with multi-round negotiation and a PPO-based election actor, and RadioDynamics evaluates it across 62 metropolitan scenes and 11 testing scenes.The environment integrates UAV mobility, radio propagation, inter-UAV communication modeling, and digital-twin geometry with ray-traced fields; WONDER reaches 0.870 and 100% gateway connectivity.

APPENDIX A COMMUNICATION PAYLOAD CALCULATION … B. Election Module Objective

The appendices quantify communication payloads for local aggregation and WONDER proposals, then specify JEPA world-model and PPO election objectives. The election reward combines balanced-objective improvement with a gateway-connectivity penalty, while termination receives zero reward.

  • A. Communication Accounting for Local Vectors and WONDER Proposals: 24 directed transmissions arise per local exchange from audited subgraphs of sizes 3, 4, and 3, with synchronization aggregation factor 5 × 24.The factor applies to methods aggregating over the local topology.
  • A. Communication Accounting for Local Vectors and WONDER Proposals: WONDER’s payload count is 3×[(10×3)+(3×3×2)]×32 float32 values, while STACCA, MAGI, and Full Connection use their specified local-vector or RSRP-map counts.STACCA and MAGI use 32-float and 64-float messages, respectively; Full Connection transfers a 350 × 350 RSRP map.
  • A. JEPA Radio World Model Objectives: The JEPA radio world model includes an auxiliary ranking loss for the best trajectory index k∗ of each UAV i in round r.The objective averages over active UAV-round pairs and uses a fixed positive score-normalization constant σscore.
  • B. Election Module Objective: At each election round, Aϕ selects one trajectory from the global trajectory pool or ends negotiation, updating the partial joint plan when a trajectory is committed.The deterministic reductions produce the global pool, and the committed trajectory defines the post-election plan.
  • B. Election Module Objective: The election reward measures balanced-objective improvement and penalizes gateway-connectivity violations with weight λc.The terminal metric vector is evaluated after executing the plan over the synchronization horizon.
  • B. Election Module Objective: The election module is optimized with a PPO objective combining clipped policy loss, centralized value loss, and entropy regularization.The clipped loss uses the PPO probability ratio and advantage estimate; cv and ce weight value-loss and entropy terms, respectively.
  • B. Election Module Objective: Zero transition reward is assigned when Aϕ ends negotiation instead of selecting a trajectory.The PPO action set is formed from the global trajectory pool and the termination action.
Loading 2608.16955v1…