Camillo Gentile

Data is becoming an increasingly essential part of wireless communications research. It is used for training machine learning models, testing novel resource management and allocation algorithms, and extracting novel insights about populations and systems. However, useful wireless networking data is often challenging and expensive to obtain. In this paper, we conduct an AI-assisted meta-study measuring wireless dataset reuse in five networking publication venues over four years. Our results, which span over 6 000 papers, show that dataset reuse is rare in our dataset, with most datasets being unavailable to the public. While there may be many reasons for this lack of dataset reuse, our finding contrasts starkly with the machine learning community, where well-known standardized benchmark datasets (e.g., MNIST, CIFAR-10, ImageNet) have catalyzed massive research efforts, allowing researchers to build upon and validate each others' work. We propose some properties that will be important for the next generation of wireless datasets to better support research, particularly in the subfield of AI for wireless. We conclude with a call to action to decision-makers and community members on how we can help to realize the creation of standardized, widely-adopted datasets for wireless research, particularly in the field of AI for wireless.

Cross-polarization discrimination (XPD) is a key parameter for characterizing dual-polarized millimeter-wave (mmWave) wireless channels, where surface interactions induce depolarization. Existing models treat XPD either statistically—by aggregating power across all paths—or deterministically—based on idealized geometry—but rarely provide path-specific predictions grounded in measurement. To bridge this gap, this paper presents a measurement-based geometric–stochastic channel model that predicts per-path XPD as a function of incidence geometry and material. We introduce the off-specular angle as a unifying geometric parameter linking incidence geometry to depolarization. Three measured quantities—path count, angular offset, and XPD—were each found to follow distinct distribution types (Poisson, Gaussian, and exponential respectively), with parameters fitted from an exhaustive 28.5 GHz polarimetric measurement campaign spanning ten material types and 266 material orientations. The fitted trends reveal how depolarization varies with material class, geometric alignment, and polarization direction. The resulting framework enables integration of material-dependent depolarization into raytracing engines and geometry-based stochastic channel models (GSCMs), supporting more realistic simulation of polarization effects in mmWave propagation.

Precisely modeling radio propagation in complex environments has been a significant challenge, especially with the advent of 5G and beyond networks, where managing massive antenna arrays demands more detailed information. Traditional methods, such as empirical models and ray tracing, often fall short, either due to insufficient details or because of challenges for real-time applications. Inspired by the newly proposed 3D Gaussian Splatting method in the computer vision domain, which outperforms other methods in reconstructing optical radiance fields, we propose RF-3DGS, a novel approach that enables precise site-specific reconstruction of radio radiance fields from sparse samples. RF-3DGS offers high efficiency, requiring only a few minutes for training and achieving fast inference for any arbitrary receiver pose within milliseconds. Furthermore, RF-3DGS can provide fine-grained Spatial Channel State Information (Spatial-CSI) of these paths, including the channel gain, the delay, the angle of arrival (AoA), and the angle of departure (AoD). Our experiments, calibrated through real-world measurements, demonstrate that RF-3DGS not only significantly improves reconstruction quality, training efficiency, and rendering speed compared to state-of-the-art methods, but also holds great potential for supporting wireless communication and advanced applications such as Integrated Sensing and Communication (ISAC). Code and dataset are available at https://github.com/SunLab-UGA/RF-3DGS

The capability to sense and classify targets through radar instead of optical sensors offers the advantage of non-line-of-sight operation, low battery consumption, and privacy, placing integrated sensing and communications at the core of 6G networks. Thanks to their fine spatial resolution, commercial millimeter-wave radios have already been repurposed for target classification–in particular for human gesturing, the topic of this paper–through deep learning trained directly on radar data as input. While promising, to achieve robust classification across a wide variety of humans and scenes, thousands to tens of thousands of training exemplars are needed. Such databases are unavailable to date, and developing application-specific databases may require prohibitive durations, from months to years. To alleviate the burden of training, we integrate keypoints–a set of 3D points tagged to the salient body parts, generated from a camera/lidar system as an add-on to our in-house 28 GHz radar system–as an intermediate step of the training process, between the radar input and the classification output. The keypoints exploit rich databases of millions of exemplars of different humans and scenes, that are publicly available. By integrating keypoints, we can achieve 90 % classification accuracy with fewer than 100 exemplars. In addition, we can operate at distances up to 2.5 m versus the near-field operation (within 50 cm) of most other systems and we are the first to consider 16 gesture classes across both forearms rather than just a few classes across a single forearm.

Millimeter-wave (mmWave) communications are a key enabler of next-generation wireless networks, thanks to their abundant spectrum resources. Cluster-based models offer a compact and physically interpretable abstraction by grouping multipath components (MPCs) with similar properties, but their accuracy and adaptability are often limited by insufficient environmental information. To address this limitations, we propose a predictive channel modeling framework that jointly estimates MPCs clusters and associated parameters by leveraging multimodal environmental data, including light detection and ranging (LiDAR), images, and locations. The framework extracts modality-specific features and employs a hierarchical attention-based fusion architecture to integrate complementary information. Furthermore, a transfer learning strategy is introduced to support efficient model adaptation across different indoor regions. Experimental results demonstrate that multimodal fusion significantly outperforms single-modality baselines, which reduces a relative root mean squared error (RMSE) of RMS delay spread (DS) prediction by over 50% and improves cluster classification accuracy by around 6%. Meanwhile, the multi-task learning brings an additional about 5% gain in cluster prediction. Moreover, in inter-region transfer scenarios, the model achieves over 87% accuracy using only 20% data, highlighting strong data efficiency and adaptability.

With the development of Integrated Sensing and Communication (ISAC) for Sixth-Generation (6G) wireless systems, contactless human recognition has emerged as one of the key application scenarios. Since human gesture motion induces subtle and random variations in wireless multipath propagation, how to accurately model human gesture channels has become a crucial issue for the design and validation of ISAC systems. To this end, this paper proposes a deep learning-based human gesture channel modeling framework for ISAC scenarios, in which the human body is decomposed into multiple body parts, and the mapping between human gestures and their corresponding multipath characteristics is learned from real-world measurements. Specifically, a Poisson neural network is employed to predict the number of Multi-Path Components (MPCs) for each human body part, while Conditional Variational Auto-Encoders (C-VAEs) are reused to generate the scattering points, which are further used to reconstruct continuous channel impulse responses and micro-Doppler signatures. Simulation results demonstrate that the proposed method achieves high accuracy and generalization across different gestures and subjects, providing an interpretable approach for data augmentation and the evaluation of gesture-based ISAC systems.

Clustering radio-frequency (RF) multipath components (MPCs) fosters compact channel models by capturing the geometry of the scattering environment, yet “blind” methods based solely on RF data struggle to associate MPCs with individual scatterers. We address this by supplementing the RF channel sounder with camera and lidar systems that, together with AI-based algorithms, generate a segmented digital twin of the environment. By projecting MPCs onto this segmented twin, they are clustered directly according to the segments they fall on. We validate the framework on a robotic arm in a joint communications and sensing (JCAS) scenario, demonstrating reliable clustering by individual robot parts.