RadioDreamer: A World Model for Radio Map Forecasting and Imagination-Based Optimization

I. INTRODUCTION

• Existing learning-based methods map scenes and fixed network configurations to radio-map snapshots rather than modeling action-driven evolution.RadioUNet and PMNet regress path-loss distributions, while RadioDiff uses conditional generation.Source: §I · p. 1, §II · p. 2

II. RELATED WORK

• World models learn action-conditioned dynamics and enable imagined trajectories for evaluating long-term effects before execution.Prior applications cited include embodied intelligence,, robotic control, and autonomous driving.Source: §I · p. 1

I. INTRODUCTION

• RadioDreamer learns action-conditioned latent dynamics from sequential observations, then rolls out candidate actions to optimize long-term network configurations.The paper describes it as the first world model for action-conditioned radio-map forecasting and network optimization in cellular networks.Source: §I · pp. 1–2
• Its Cellular-Aware Transformer Block models temporal dependencies, inter-BS interactions, and spatial propagation; Region-Adaptive Supervision emphasizes dynamic, action-sensitive regions.The supervision is reported to improve localized prediction and reduce multi-step rollout error accumulation.Source: §I · p. 2
• 25.2% average reduction in full-map forecasting error and 8.8% higher episode return over RM-based optimization are reported, alongside up to 45.2% lower dynamic-region error.Evaluations span three urban scenes and three generalization protocols.Source: §I · p. 2

II. RELATED WORK

• RadioDreamer addresses the gap between static radio-map reconstruction and the need to compare candidate actions’ long-term consequences.Prior radio-map work includes electromagnetic-knowledge methods and foundation models [21]–, while the paper frames its contribution as dynamic forecasting and counterfactual optimization.Source: §II · p. 2

III. PROBLEM FORMULATION

• The environment is an R × C grid with static buildings, base stations, and moving objects whose motion affects propagation paths.The static scene includes building geometry, material properties, and base-station locations.Source: §III · p. 2
• Radio maps characterize RSRP and SINR distributions and evolve jointly with base-station actions and environmental mobility.Base-station reconfiguration changes signal strength, coverage boundaries, and inter-cell interference; moving objects add time-varying propagation effects.Source: §III · p. 2
• The forecasting objective is to predict future radio maps from historical maps, base-station actions, moving-object states, the static scene, and future base-station actions.The model represents the conditional distribution of future maps given historical context and future actions.Source: §III · p. 2
• Future moving-object states are unavailable and are inferred from the historical sequence; heterogeneous interactions make recursive long-horizon prediction challenging.The challenge spans spatial, temporal, and base-station dimensions, with inadequate dynamics modeling risking rollout error accumulation.Source: §III · p. 2

A. RadioDreamer Paradigm

• RadioDreamer learns action-conditioned radio-environment dynamics in Stage I, then uses the learned latent dynamics to evaluate and optimize BS actions in Stage II.Source: A. RadioDreamer Paradigm · p. 3

B. Stage I: Dynamics Learning

• Four modality-specific tokenizers encode scene and BS topology, moving-object states, historical radio maps, and historical and future BS actions before fusion.Source: B. Stage I: Dynamics Learning · p. 3
• The Cellular-Aware Transformer Block factorizes interactions across time, BSs, and the row and column dimensions of the radio-map grid.Causal attention prevents latent states from accessing future radio-state tokens.Source: B. Stage I: Dynamics Learning · pp. 3–4
• Region-Adaptive Supervision uses sample- and time-specific dynamic masks to emphasize meaningful radio-map variations rather than predefined regions.The masks target dynamics associated with moving objects, coverage transitions, and interference fluctuations.Source: B. Stage I: Dynamics Learning · p. 4
• A residual decoder predicts future radio-map changes relative to the latest observed map, retaining its propagation background as a common reference.Source: B. Stage I: Dynamics Learning · p. 5
• The dynamics objective combines global prediction with dynamic-region supervision to preserve propagation features while modeling changes from BS reconfiguration and moving objects.The dynamic-region loss compares ground-truth and predicted variations within selected regions.Source: B. Stage I: Dynamics Learning · pp. 4–5

• The Planner selects BS actions in a closed loop, while the frozen dynamics model predicts latent states and a Reward Model scores imagined state–action pairs.This produces multi-step imagined trajectories without decoding each intermediate state into a full radio map.Source: C. Imagination Planning · p. 5
• Sampling alternative power and downtilt actions from the same latent state generates counterfactual trajectories for long-term policy optimization without physical-network interaction.Source: C. Imagination Planning · p. 5
• During policy learning, the Critic estimates long-term return from imagined latent states and the Actor selects BS actions to maximize cumulative imagined reward.The Actor retains an entropy term to encourage exploration among alternative configurations.Source: C. Imagination Planning · pp. 5–6

• The evaluation asks whether RadioDreamer reliably predicts radio-map evolution in dynamic radio environments.Source: §V · p. 6
• It tests whether RadioDreamer models action-state dynamics and generalizes to unseen BS configurations.Source: §V · p. 6
• It assesses the effectiveness of imagination-based planning for BS-control optimization.Source: §V · p. 6

• Each 32-frame episode contains 100 moving objects and 2–6 active BSs, with transmit powers from 10 to 30 dBm and downtilts selected from five listed values.Training, validation, and test trajectory seeds are mutually exclusive.Source: A. Evaluation Setup · p. 6
• The dataset contains 483,840 temporally ordered radio-environment frames across three independently trained urban scenes, generated with Sionna RT.Each scene uses a 48×32 map at 2.5 m spatial resolution over a 120 m×80 m region.Source: A. Evaluation Setup · p. 6
• T1 holds out trajectory instances, T2 holds out BS coordinates, and T3 holds out both BS coordinates and mobility regimes.Source: A. Evaluation Setup · p. 6
• Baselines include a 3GPP statistical model, five static radio-map models, and Dreamer-RSSM as a general-purpose world model.Source: A. Evaluation Setup · p. 6
• Forecasting is evaluated with F-MAE over the full map, D-MAE over masked dynamic regions, and Signal Direction accuracy for radio-map variation.Source: A. Evaluation Setup · pp. 6–7

B. Forecasting performance (RQ1)

• 25.2% average F-MAE reduction over the strongest baseline, with RadioDreamer lowest in all 18 scene, radio-quantity, and horizon combinations.At H = 16 in Munich, it reduces RSRP and SINR errors by 32.1% and 35.4%, respectively, versus Dreamer-RSSM.Source: B. Forecasting performance (RQ1) · p. 7
• 6.5%–45.2% lower dynamic-region MAE than the strongest competitor, with approximately 80% SD versus 48.07%–53.60% for the best baselines.The results indicate that low full-map error alone does not ensure accurate prediction of local radio-change direction.Source: B. Forecasting performance (RQ1) · p. 7
• RadioDreamer’s ∆-MAE is 0.93–1.13 dB, versus 1.55–2.07 dB for Pathloss-PINN and 1.61–1.88 dB for PMNet.The qualitative comparison reports closer matching to Sionna RT for both positive and negative SINR changes; Pathloss-PINN misses localized changes, while PMNet retains detail but predicts inaccurate variation patterns.Source: B. Forecasting performance (RQ1) · p. 7, Figure 4 · p. 7

C. Action-state Dynamics and Generalization (RQ2)

• 43.5% and 46.7% lower MAEs than the strongest baselines for Power–RSRP and Downtilt–SINR responses, respectively.At H = 8, the reported MAEs are 0.013 dB and 0.008 dB, and the model most closely matches the Sionna RT dose–response curves.Source: C. Action-state Dynamics and… · pp. 7–8, Figure 5 · p. 8
• RadioDreamer captures the direction, magnitude, and nonlinear response of radio-state changes induced by BS actions, while its structured dynamics modeling supports generalization beyond training configurations.The reported explanation highlights temporal action–state dependencies, inter-BS interactions, and spatial propagation.Source: C. Action-state Dynamics and… · p. 8
• 9.6%–26.9% lower F-MAE than Dreamer-RSSM across all six T2 and T3 settings, with an average reduction of 20.3%.In Munich, RadioDreamer achieves 1.943 dB under T2 and 2.029 dB under T3; comparable T2/T3 performance is reported for jointly held-out BS positions and unseen trajectories.Source: C. Action-state Dynamics and… · p. 8, Table III · p. 8

D. Imagination Optimization (RQ3)

• The task optimizes a 16-step sequence of transmit-power and electrical-downtilt actions for every active base station in a scene with 100 moving objects and 2–6 active base stations.The objective addresses evolving coverage and interference rather than a single radio-map snapshot.Source: D. Imagination Optimization (RQ3) · p. 8
• The reward combines broad coverage and weak-region coverage, using RSRP ≥ −100 dBm and SINR ≥ 0 dB thresholds and a reference region containing the lowest 5% of RSRP values.Cell-edge coverage is the proportion of that reference region whose RSRP exceeds −55 dBm.Source: D. Imagination Optimization (RQ3) · p. 8
• 76.88 ± 2.26 episode return, the highest reported, exceeding RM-based SAC by 8.8% and decoded-radio-map planning by 2.0%.It also improves over the default configuration and one-step greedy optimization by 22.5% and 36.1%, respectively.Source: D. Imagination Optimization (RQ3) · p. 9, Figure 6 · p. 9
• 95.05% RSRP coverage, 71.45% SINR coverage, and 41.90% 5%-worst-region coverage, ranking first on all three metrics.Compared with the default configuration, these values increase by 1.58, 40.01, and 23.55 percentage points, respectively.Source: D. Imagination Optimization (RQ3) · p. 9, Table IV · p. 9

E. Ablation and model complexity

• Without temporal causal attention, H = 16 RSRP/SINR F-MAE rises from 0.893/1.066 to 1.313/1.656, and D-MAE rises from 2.584 to 5.118.The study identifies temporal causal attention as the largest ablation contribution.Source: E. Ablation and model complexity · p. 9, Table V · p. 9
• Without Region-Adaptive Supervision, H = 16 RSRP/SINR F-MAE rises to 1.131/1.349 and D-MAE nearly doubles to 5.104.Removing BS-level or spatial attention also degrades performance, and excluding historical motion information raises D-MAE to 3.197.Source: E. Ablation and model complexity · p. 9
• Approximately 10.5 ms per inference, with 112 MiB peak GPU memory; latency is about half that of Dreamer-RSSM.The reported comparison describes a favorable balance between forecasting accuracy and computational cost.Source: E. Ablation and model complexity · p. 9, Figure 7 · p. 9

VI. CONCLUSION

• 25.2% lower radio-map forecasting error on average than the strongest baselines across three urban scenes.Source: §VI · p. 9
• 8.8% higher episode return than planning over decoded radio maps for network optimization.Source: §VI · p. 9
• Structured attention and region-adaptive supervision support informative latent representations and multi-step imagination for network optimization.Source: §VI · p. 9
• Future work will add real-world measurements, unified training across urban scenes, and broader network planning and optimization tasks.Source: §VI · p. 9