"Multi-Configuration FMCW Radar Dataset for Hand Gesture Recognition"

"Frequency-modulated continuous wave (FMCW) radar offers a highly robust, privacy-preserving alternative to traditional optical sensors for hand gesture recognition. However, a major bottleneck in deploying radar-based neural networks is their sensitivity to sensor configuration shifts. Models trained on a fixed set of radar parameters frequently suffer from severe performance degradation when real-world application requirements necessitate changes to bandwidth, sampling rate, or chirp timing.To bridge this gap and drive research in configuration-agnostic learning, we introduce a comprehensive, large-scale multi-configuration radar dataset. This benchmark contains 72,000 independent gesture samples, capturing a diverse range of signal distributions. The data was recorded across 24 distinct radar configurations. Furthermore, to ensure hardware diversity and account for slight manufacturing variations, the dataset was collected using three separate physical Infineon XENSIV\u2122 BGT60TR13C 60GHz FMCW radar sensors.We believe this dataset will be an invaluable resource for the machine learning and signal processing communities. By providing explicit, systematic distributional shifts, we hope to accelerate the development of continual learning mechanisms, universal radar foundation models, and adaptive human-computer interaction systems that operate seamlessly regardless of the underlying radar setup. This dataset is a contribution of our paper submitted to the European Radar Conference (EuRAD) at the European Microwave Week 2026. To the best of our knowledge, it represents the largest open-source multi-configuration radar dataset available to date."

调频连续波雷达（Frequency-modulated continuous wave, FMCW）在手势识别任务中，相较于传统光学传感器，是一类鲁棒性极强、可保护用户隐私的替代方案。然而，基于雷达的神经网络在部署时面临一大核心瓶颈：其对传感器配置的变化极为敏感。当实际应用场景需要调整带宽、采样率或啁啾时序时，基于固定雷达参数训练的模型往往会出现严重的性能衰减。

为填补这一研究空白、推动配置无关学习领域的发展，我们构建了一个大规模、多配置的综合性雷达数据集。该基准数据集包含72000个独立手势样本，覆盖了多样化的信号分布场景，且采集自24种不同的雷达配置方案。此外，为保证硬件多样性并覆盖微小的制造工艺差异，本次数据集采集使用了三台独立的英飞凌XENSIV™ BGT60TR13C 60GHz调频连续波雷达传感器。

我们相信，该数据集将为机器学习与信号处理领域的研究者提供极为宝贵的研究资源。通过引入明确且系统化的分布偏移场景，我们期望能够加速持续学习机制、通用雷达基础模型以及自适应人机交互系统的研发进程——此类系统可在无需适配底层雷达配置的前提下实现无缝运行。

本数据集源自我们提交至2026年欧洲微波周旗下欧洲雷达会议（EuRAD）的学术论文。据我们所知，这是目前公开可用的规模最大的多配置雷达数据集。