mobile mmwave sensing

To address privacy leakage, illumination sensitivity, and poor occlusion robustness in existing camera-based gesture recognition systems, this paper proposes an end-to-end, non-contact human–robot interaction system leveraging millimeter-wave (mmWave) radar. The system uniquely integrates radar-based gesture recognition with robot control via a behavior tree framework, enabling real-time mapping of nine distinct gestures to robotic arm commands. It achieves high-accuracy recognition and low-latency response under challenging conditions—including varying illumination and partial occlusions. Key technical contributions include real-time radar signal processing, a lightweight deep classification model, and a behavior-tree-driven dynamic command scheduling mechanism. Experimental evaluation demonstrates an average gesture recognition accuracy of 98.2% and an end-to-end system latency below 120 ms, significantly improving interaction continuity, environmental adaptability, and overall system robustness.

This work addresses the challenge of constructing accurate probabilistic occupancy maps in adverse environments such as smoke or dense fog, where conventional sensors often fail. While millimeter-wave radar offers robustness under such conditions, its signals are typically sparse and noisy, hindering high-fidelity mapping. To overcome this limitation, the paper presents the first end-to-end framework that integrates synthetic aperture radar (SAR) processing with probabilistic occupancy mapping. The authors systematically analyze the impact of antenna array configurations and key system parameters on mapping performance. Leveraging GPU acceleration for efficient computation, the proposed method is validated across multiple indoor scenarios, with map quality quantitatively assessed through path planning tasks. Additionally, the study introduces and publicly releases the first cascaded millimeter-wave radar dataset along with a complete open-source toolchain.

This work addresses the challenge of dynamic obstacle avoidance for small quadrotors during high-speed flight, particularly under adverse conditions such as abrupt illumination changes or smoke. It presents the first onboard, real-time collision avoidance system based on millimeter-wave radar. The proposed framework integrates a lightweight interacting multiple model (IMM) tracker with a control barrier function (CBF)-based controller in a perception-control co-design architecture that directly outputs collision-avoidance accelerations satisfying spatiotemporal constraints. The authors theoretically derive sufficient conditions guaranteeing successful avoidance. Experimental results demonstrate centimeter-level accuracy across 390 trials, with position errors below 0.15 m, 0.93 m, and 0.87 m in the x, y, and z axes, respectively, and an end-to-end latency of approximately 14 ms, enabling robust operation in both bright/dark transitions and smoky environments.

This work addresses the challenge of irreproducibility in indoor millimeter-wave radar sensing, which arises from scene geometry, material properties, multipath effects, and hardware or signal processing discrepancies. To overcome this, we present the first signal-level, path-traceable digital twin platform for indoor millimeter-wave radar. By integrating Unreal Engine–based ray tracing with FMCW signal modeling, the platform generates complex-valued multi-channel received signal grids along with per-path contribution records. It shares common interfaces for receive channels and range-angle processing, enabling direct comparison between simulated and real-world measurements. The system supports path-level provenance across objects, materials, propagation events, and output intervals. Evaluated in an office environment across 22 locations and 154 poses, it recalls 70.8% of measured active geometric response regions within the central field of view, effectively revealing residual sources and establishing a diagnosable, reproducible workflow for practical deployment.

This study addresses the challenge of accurate terrain perception for agricultural drones operating in complex farmland environments, where occlusions, elevation variations, and environmental disturbances often degrade low-altitude flight performance. To overcome these limitations, the authors propose a terrain-awareness framework based on a low-cost rotating millimeter-wave radar. By mechanically rotating the radar to expand its field of view, they develop a pose-consistent pipeline for sparse radar point cloud registration and filtering, followed by a novel ground segmentation and surface reconstruction algorithm tailored for high-noise, partially observable data. As the first work to apply a rotating millimeter-wave radar to terrain perception in agricultural drones, the method achieves a ground segmentation F1-score of 94.42 in real-world experiments—significantly outperforming existing approaches (90.48)—and demonstrably enhances terrain coverage, estimation accuracy, and flight robustness.

This work addresses the challenge of limited generalization in millimeter-wave radar perception due to scarce real-world data across novel objects, environments, and trajectories. The authors propose a scene-customized simulation framework that requires no real radar data: leveraging 3D reconstructions of target scenes and surface material inference via vision-language models, it generates high-fidelity FMCW radar signals through physically accurate ray tracing and multipath modeling. This approach enables, for the first time, the synthesis of realistic radar training data prior to deployment, demonstrating strong correspondence between simulated and real radar responses in terms of shape and material characteristics. Using only synthetic data, the method achieves an object recognition accuracy 2.5 times higher than random guessing; with minimal real labeled data, it attains 95.3% accuracy on a 12-class object recognition task.

This study addresses the challenge of unstable respiratory and cardiac monitoring using millimeter-wave radar in unconstrained home environments, where performance degrades due to sensitivity to observation geometry. The authors propose a vision-radar cooperative active perception framework that, for the first time, incorporates sensing geometry as a controllable variable in robotic motion planning. By leveraging visual guidance to dynamically adjust the radar’s pose—aligning it nearly perpendicularly to the thoracic surface—the system maximizes observability of radial physiological motion. This approach establishes a closed-loop perception-action pipeline integrating visual keypoint localization, robotic motion control, millimeter-wave signal processing, and differential phase enhancement. Experimental results demonstrate significant improvements: respiratory interval error decreases from 0.87 s to 0.14 s, and heart rate error drops from 13.59 bpm to 2.22 bpm, achieving accuracy in free-living conditions comparable to that in static, controlled settings.

本文综述了4D毫米波雷达感知算法在自动驾驶中的应用，从信号处理到动态场景重建，比较了不同方法，并讨论了未来研究方向。

This work addresses the limitation of existing 4D millimeter-wave radar modeling and scan-matching approaches, which often neglect radar cross-section (RCS) information, leading to inadequate scene representation. To overcome this, the authors propose a 4D radar Gaussian model that explicitly incorporates RCS as a physical attribute alongside conventional position and Doppler velocity measurements. This is the first method to integrate RCS into both the 4D Gaussian representation and the scan-matching pipeline. By embedding RCS into the geometric and kinematic description of radar points, the proposed approach enhances the semantic expressiveness of the point cloud and significantly improves the accuracy and robustness of registration. Consequently, it achieves superior radar-based mapping performance in complex environments.

Millimeter-wave radar signals are challenging to align with natural language, and existing approaches either rely on synthetic data or lack explicit supervision of human body structure and motion, limiting their effectiveness in behavior understanding. This work proposes mmMind, which introduces synchronized 3D human pose as a supervisory signal exclusively during training to pretrain a spatiotemporal radar encoder for learning human configuration and dynamics. At inference, the model operates solely on radar input and generates behavior descriptions and answers spatiotemporal questions by aligning with a large language model. The study establishes mmMind-Bench, the first real-world millimeter-wave radar–language benchmark, and demonstrates significant performance gains over current baselines across multiple tasks, validating the efficacy and generalization capability of pose-guided pretraining.