A reconfigurable intelligent surface is a flat panel of many small elements that redirect a radio signal by shifting its phase. The panel adds a bounce path around walls and machines that block high-frequency links. Mounting such a panel on a robot or drone lets it follow users and keep that bounce path open, but the controller that decides where to drive and how to set the phases has to be learned with reinforcement learning, separately for each building or floor. Prior work speeds this up with federated learning, in which each site's controller trains locally and only the learned weights are averaged. It averages every site equally, however, even when one floor is an open office and another is a dense machine hall. This proposal asks a simple question: can each site's environment be summarized before any training, using the standardized texture statistics that medical imaging calls radiomics, so that each controller averages mainly with sites that resemble it?
The idea in two sentences is this. Treat each site's radio coverage map as an image, extract the same few hundred texture and intensity features that radiologists extract from CT scans, and compare sites with the Fréchet radiomic distance, a distribution distance designed to stay reliable with as few as ten images. Then use that distance to weight the federated averaging of the mobile-surface controllers, and to warm-start new sites from their nearest neighbors, and test whether this beats both plain averaging and similarity measures that need training to compute.
Core Claim
The claim is that environment heterogeneity, not compute, is what limits federated training of mobile-surface controllers once sites differ in layout. It also holds that a small-sample, interpretable distribution distance over radio-map textures predicts which sites' policies transfer to each other. Prediction from this distance is available at round zero. Gradient-based client clustering cannot match that, because it needs training rounds before it can tell sites apart.
Two papers motivate the setting. Mobile Reconfigurable Intelligent Surfaces for NOMA Networks: Federated Learning Approaches puts surfaces on indoor robots. It trains a deep deterministic policy gradient controller per office floor to jointly choose robot position, phase shifts and access-point power, and it federates those controllers with uniform averaging. It reports peak reward in 150 episodes versus 250 for independent training, a 40% reduction, plus a 15.1% rate gain over an optimally placed static surface. Its floors come from a single office propagation model, and it never asks whether averaging hurts when floors differ. The general warning that non-identically distributed clients pull local training away from the global objective comes from Advances and Open Challenges in Federated Foundation Models. We test that warning directly in the control setting.
The method comes from Fréchet radiomic distance (FRD) (see also the companion record). FRD extracts 464 standardized first-order, texture-matrix, wavelet and Laplacian-of-Gaussian features. It reports an out-of-distribution detection AUC of 0.94 against 0.79 for the deep-feature Fréchet Inception Distance, and stability at sample sizes of 10 to 100 where the deep metric diverges below about 1,000. That small-sample property is the reason to borrow it: a newly deployed robot has only a handful of coverage snapshots. Brain Tumor Survival Prediction using Radiomics Features gives independent support that hand-built features beat a learned network at small sample sizes, with 76.5% versus 51.5% accuracy on 166 BraTS patients.
The distance between sites $i$ and $j$ is
$$ d_{ij} = \lVert \mu_i - \mu_j \rVert^2 + \mathrm{Tr}\!\left(\Sigma_i + \Sigma_j - 2(\Sigma_i \Sigma_j)^{1/2}\right), $$
where $\mu_i$ and $\Sigma_i$ are the mean and shrinkage-regularized covariance of the radiomic feature vectors computed from site $i$'s coverage-map snapshots. Site $i$ then receives the personalized average
$$ w_i^{t+1} = \sum_j \alpha_{ij} w_j^{t}, \qquad \alpha_{ij} = \frac{\exp(-d_{ij}/\tau)}{\sum_k \exp(-d_{ik}/\tau)}, $$
where $w_j^t$ is site $j$'s controller weights at round $t$ and $\tau$ is a temperature. When $\tau$ is very large, every site gets equal weight and this reduces to Zhong et al.'s uniform averaging. When $\tau$ approaches zero, each site trains alone. Sites share only $\mu_i$ and $\Sigma_i$, never maps, floor plans or trajectories. The medical-radiomics literature also supplies a constraint. Making head and neck cancer clinical data FAIR found that radiomic features added nothing over clinical variables across five federated cohorts, with concordance 0.60–0.72 versus 0.62–0.73. A plausible cause is that features shift with acquisition settings. The radio-map analogue of that shift is grid resolution, power quantization and map extent, so all sites must compute maps on a fixed physical grid and power scale. We treat that standardization as a design requirement and stress-test it, rather than assuming it.
Research Program
The first step is a heterogeneous environment suite. We will reimplement the indoor office propagation model of Zhong et al. (ITU-R P.1238 path loss, randomly placed blocking "bricks", Rician fading, roaming users) and vary layout systematically: blockage density, corridor versus open-plan structure, clustered machinery typical of the robotic factories surveyed in Reconfigurable Intelligent Surface for Internet of Robotic Things, and access-point placement. This yields about 40 distinct sites. As a second, harder family, we use 3GPP urban street-canyon blocks with drone-mounted surfaces, following Reflections in the Sky. That paper reports line-of-sight probability above 90% for the aerial surface against below 5% for a static one. The drone family checks whether the method survives 3D motion and propulsion-energy terms of the kind handled by DRL in Energy-Efficient Communication Networks via Multiple Aerial RIS and Aerial STAR-RIS Empowered MEC.
The second step builds a ground-truth transfer matrix. For every pair of sites, we train a controller in site $i$, evaluate it zero-shot and after short fine-tuning in site $j$, and record the normalized loss in reward. This matrix is the target any environment-similarity measure should predict, and it is expensive. That cost is the reason a cheap proxy is worth having.
The third step computes each site's radiomic fingerprint. Inputs are snapshots of two channels on the standardized grid: the received-power map without the surface, and the blockage mask. Snapshots are taken as users and transient blockers move. We then measure how well $d_{ij}$ predicts the transfer matrix, as a function of the number of snapshots from 5 to 200.
The fourth step closes the loop. Similarity-weighted federated training runs alongside a warm-start protocol: a newly arriving site, which has only a few snapshots and no trained weights, initializes from the $\alpha$-weighted average of existing controllers. We also report which feature families drive the distance, for example run-length non-uniformity of the blockage mask versus first-order statistics of received power. This keeps the result interpretable to wireless engineers.
Baselines
The naive reference is independent training: each site learns its own controller with no sharing. Zhong et al. report this needs 250 episodes to peak in their setting. We rerun it on our suite with the same network, exploration schedule and episode budget as every other method.
The strongest existing method is uniform federated averaging of the DDPG controllers as in Mobile Reconfigurable Intelligent Surfaces for NOMA Networks, which reports peak reward at 150 episodes (a 40% reduction) and a 15.1% rate gain over the best static surface. Those numbers come from a fairly homogeneous office family. We will first reproduce them on our homogeneous subset to validate the simulator, then run on the heterogeneous suite with identical hyperparameter search budgets for $\tau$-free FedAvg and our method. Beating it means one of two things on heterogeneous sites: reaching 95% of each site's own peak reward in significantly fewer episodes (paired across ten seeds), or reaching a higher final sum rate. Matching it on the homogeneous subset is the sanity requirement, since the method reduces to FedAvg there.
The competitor most likely to tie us is clustered or personalized federated averaging that weights sites by similarity of their parameter updates. It has no domain-specific assumptions, and on this specific mobile-surface task there is no reported result found on Lacuna. We implement it with the same softmax weighting, swapping $d_{ij}$ for cosine distance between update vectors, and tune its temperature with the same budget.
The upper bound is an oracle that sets $\alpha_{ij}$ from the measured transfer matrix. Three ablations isolate the key ingredient, the texture distribution distance. The first replaces $d_{ij}$ with scalar descriptors: blockage density and mean path loss. The second replaces radiomic features with embeddings from a convolutional autoencoder trained on all maps, which is a Fréchet Inception-style distance. The third drops the covariance term and keeps only the mean difference.
Why not just…?
Why not just average everything, as Zhong et al. already do? On near-identical floors, uniform averaging is probably close to optimal, and our method is built to collapse to it there. The open question is heterogeneous sites, where averaging an open-plan policy into a machine-hall policy can steer the robot toward positions that are blocked in one of them. The transfer-matrix experiment shows directly whether such negative transfer exists. If off-diagonal transfer losses are uniformly small, the premise fails, and we would report that federated mobile-surface control does not need personalization.
Why not just cluster sites by gradient similarity? This is the real threat. After enough rounds it may match us, because gradients encode environment differences implicitly. Our claim is limited to what it cannot do: weight sites at round zero, and warm-start a new site before it has trained. The warm-start experiment settles this. If gradient clustering, given the same few episodes a new site would spend collecting snapshots, matches our zero-shot initialization, the radiomic fingerprint offers no advantage and we will say so.
Why not just use one or two hand-picked numbers like blockage density? That would be far simpler and more convincing to a wireless reviewer. It is the first ablation. If scalar descriptors predict the transfer matrix as well as the full feature set (compared by Spearman correlation with bootstrapped intervals), the texture features are unnecessary decoration. The radiomic contribution then shrinks to a negative result.
Why not apply mobile surfaces to carry federated radiomics models between hospitals instead? This is the obvious intersection, and we rejected it. Federated radiomic models such as the Cox regressions in Making head and neck cancer clinical data FAIR have at most a few hundred coefficients, so the wireless uplink is not the bottleneck any ordinary link would struggle with. That paper's failure, where radiomics added nothing over clinical variables, points to cross-site feature shift, which a better radio link cannot fix. The over-the-air aggregation designs in RIS Enabled Federated Learning and RIS Empowered Over-the-Air Federated Edge Learning are aimed at large neural updates from many devices, not tiny clinical models.
Proposed Experiments
The first decisive check is predictive. Across all site pairs in both environment families, we compute the Spearman correlation between $d_{ij}$ and the measured transfer loss, alongside correlations for scalar descriptors, autoencoder distance and gradient similarity at rounds 0, 10 and 50. We plot each against the number of snapshots. The method is worth pursuing only if radiomic distance at about 10 to 20 snapshots beats scalar descriptors and approaches gradient similarity at round 50.
The second check is end-to-end training. Under matched tuning budgets and ten seeds, we compare episodes-to-95%-peak, final sum rate, and rate gain over an optimally placed static surface against independent training, uniform averaging and gradient-weighted averaging, with the oracle as the ceiling. The success threshold is a paired-significant improvement over uniform averaging on the heterogeneous suite, together with no loss on the homogeneous subset. Our reproduction of the 150-versus-250-episode result anchors that comparison.
The third check is the new-site warm start: reward over the first 20 episodes for a held-out site initialized by radiomic weighting, uniform averaging, or random initialization.
The fourth check is robustness. Deliberately mismatched map grids or power scales between sites test whether the acquisition sensitivity seen in clinical radiomics reappears, and whether the fixed-grid standardization removes it. The drone street-canyon family tests whether the result holds beyond flat indoor floors.
References
- Mobile Reconfigurable Intelligent Surfaces for NOMA Networks: Federated Learning Approaches
- Advances and Open Challenges in Federated Foundation Models
- Fréchet radiomic distance (FRD): A versatile metric for comparing medical imaging datasets
- Brain Tumor Survival Prediction Using Radiomics Features
- Making head and neck cancer clinical data Findable-Accessible-Interoperable-Reusable to support multi-institutional collaboration and federated learning
- Reconfigurable Intelligent Surface for Internet of Robotic Things
- Reflections in the Sky: Millimeter Wave Communication with UAV-Carried Intelligent Reflectors
- Energy-Efficient Communication Networks via Multiple Aerial Reconfigurable Intelligent Surfaces: DRL and Optimization Approach
- Aerial STAR-RIS Empowered MEC: A DRL Approach for Energy Minimization
- Reconfigurable Intelligent Surface Enabled Federated Learning: A Unified Communication-Learning Design Approach
- Reconfigurable Intelligent Surface Empowered Over-the-Air Federated Edge Learning