**RadarOcc: Robust 3D Occupancy Prediction with 4D Imaging Radar** **Authors:** Fangqiang Ding, Xiangyu Wen, Lawrence Zhu, Yiming Li, Chris Xiaoxuan Lu **Affiliations:** University of Edinburgh, NYU, UCL **Abstract:** This paper introduces RadarOcc, a novel approach for 3D occupancy prediction using 4D imaging radar sensors. Traditional methods rely on LiDAR or cameras, which are susceptible to adverse weather conditions. RadarOcc leverages the advantages of 4D radar, such as all-weather capability and detailed imaging outputs, to improve robustness. The method processes 4D radar tensors directly, preserving essential scene details. It addresses challenges like voluminous and noisy data through Doppler bins descriptors, sidelobe-aware spatial sparsification, and range-wise self-attention mechanisms. Spherical-based feature encoding and aggregation minimize interpolation errors. Benchmarking on the K-Radar dataset shows RadarOcc's state-of-the-art performance, even outperforming LiDAR and camera-based methods in adverse weather conditions. **Introduction:** 3D occupancy prediction is crucial for autonomous driving, offering comprehensive scene descriptions and strong generalizability. Current methods often use LiDAR or cameras, but these are limited by weather conditions. RadarOcc uses 4D imaging radar, which provides detailed imaging outputs and all-weather capabilities. The method reduces data volume through Doppler bins descriptors and spatial sparsification, and encodes features directly in spherical coordinates to avoid interpolation errors. Experiments on the K-Radar dataset demonstrate RadarOcc's superior performance and robustness in adverse weather conditions. **Related Work:** Previous work on 3D occupancy prediction has focused on LiDAR and cameras, while 4D radar has been underutilized. 4D radar offers advantages in adverse weather and all-weather sensing. Radar tensors (RTs) provide rich and complete measurements, but processing them efficiently is challenging due to large data sizes and noise. **Method:** RadarOcc consists of four components: data volume reduction, spherical-based feature encoding, spherical-to-Cartesian feature aggregation, and 3D occupancy decoding. It reduces data volume through Doppler bins encoding and sidelobe-aware sparsification, encodes features directly in spherical coordinates, and aggregates them using learnable voxel queries. Range-wise self-attention and deformable attention enhance feature encoding and aggregation. **Experiments:** RadarOcc is evaluated on the K-Radar dataset, showing state-of-the-art performance in radar-based 3D occupancy prediction. Ablation studies validate the effectiveness of key components. Qualitative results demonstrate its robustness in adverse weather conditions. **Conclusion:** RadarOcc leverages 4D imaging radar for robust 3D occupancy prediction, offering all-weather capabilities and competitive performance. Future work will address temporal information and point-wise annotations.**RadarOcc: Robust 3D Occupancy Prediction with 4D Imaging Radar** **Authors:** Fangqiang Ding, Xiangyu Wen, Lawrence Zhu, Yiming Li, Chris Xiaoxuan Lu **Affiliations:** University of Edinburgh, NYU, UCL **Abstract:** This paper introduces RadarOcc, a novel approach for 3D occupancy prediction using 4D imaging radar sensors. Traditional methods rely on LiDAR or cameras, which are susceptible to adverse weather conditions. RadarOcc leverages the advantages of 4D radar, such as all-weather capability and detailed imaging outputs, to improve robustness. The method processes 4D radar tensors directly, preserving essential scene details. It addresses challenges like voluminous and noisy data through Doppler bins descriptors, sidelobe-aware spatial sparsification, and range-wise self-attention mechanisms. Spherical-based feature encoding and aggregation minimize interpolation errors. Benchmarking on the K-Radar dataset shows RadarOcc's state-of-the-art performance, even outperforming LiDAR and camera-based methods in adverse weather conditions. **Introduction:** 3D occupancy prediction is crucial for autonomous driving, offering comprehensive scene descriptions and strong generalizability. Current methods often use LiDAR or cameras, but these are limited by weather conditions. RadarOcc uses 4D imaging radar, which provides detailed imaging outputs and all-weather capabilities. The method reduces data volume through Doppler bins descriptors and spatial sparsification, and encodes features directly in spherical coordinates to avoid interpolation errors. Experiments on the K-Radar dataset demonstrate RadarOcc's superior performance and robustness in adverse weather conditions. **Related Work:** Previous work on 3D occupancy prediction has focused on LiDAR and cameras, while 4D radar has been underutilized. 4D radar offers advantages in adverse weather and all-weather sensing. Radar tensors (RTs) provide rich and complete measurements, but processing them efficiently is challenging due to large data sizes and noise. **Method:** RadarOcc consists of four components: data volume reduction, spherical-based feature encoding, spherical-to-Cartesian feature aggregation, and 3D occupancy decoding. It reduces data volume through Doppler bins encoding and sidelobe-aware sparsification, encodes features directly in spherical coordinates, and aggregates them using learnable voxel queries. Range-wise self-attention and deformable attention enhance feature encoding and aggregation. **Experiments:** RadarOcc is evaluated on the K-Radar dataset, showing state-of-the-art performance in radar-based 3D occupancy prediction. Ablation studies validate the effectiveness of key components. Qualitative results demonstrate its robustness in adverse weather conditions. **Conclusion:** RadarOcc leverages 4D imaging radar for robust 3D occupancy prediction, offering all-weather capabilities and competitive performance. Future work will address temporal information and point-wise annotations.