RadarOcc: Robust 3D Occupancy Prediction with 4D Imaging Radar This paper introduces RadarOcc, a novel method for 3D occupancy prediction using 4D imaging radar data. The method addresses the limitations of traditional LiDAR and camera-based approaches by leveraging the advantages of 4D radar, which provides all-weather sensing capabilities and is less affected by adverse weather conditions. RadarOcc processes 4D radar tensors (4DRTs) directly, preserving essential scene details and avoiding the limitations of sparse radar point clouds. The method employs Doppler bins descriptors, sidelobe-aware spatial sparsification, and range-wise self-attention mechanisms to handle the challenges of 4D radar data. It also introduces a spherical-based feature encoding followed by spherical-to-Cartesian feature aggregation to minimize interpolation errors. RadarOcc is benchmarked on the K-Radar dataset and demonstrates state-of-the-art performance in radar-based 3D occupancy prediction, outperforming LiDAR and camera-based methods. The method also shows superior performance in adverse weather conditions. The contributions of this work include the introduction of RadarOcc for 4D radar-based 3D occupancy prediction, the development of a novel pipeline to handle 4DRT challenges, and extensive experiments validating the method's performance. The paper also discusses the limitations of the current approach, including the lack of temporal information modeling and the limitation to two general semantics. The results show that RadarOcc provides robust 3D occupancy prediction under various conditions, making it a promising alternative to LiDAR for autonomous driving.RadarOcc: Robust 3D Occupancy Prediction with 4D Imaging Radar This paper introduces RadarOcc, a novel method for 3D occupancy prediction using 4D imaging radar data. The method addresses the limitations of traditional LiDAR and camera-based approaches by leveraging the advantages of 4D radar, which provides all-weather sensing capabilities and is less affected by adverse weather conditions. RadarOcc processes 4D radar tensors (4DRTs) directly, preserving essential scene details and avoiding the limitations of sparse radar point clouds. The method employs Doppler bins descriptors, sidelobe-aware spatial sparsification, and range-wise self-attention mechanisms to handle the challenges of 4D radar data. It also introduces a spherical-based feature encoding followed by spherical-to-Cartesian feature aggregation to minimize interpolation errors. RadarOcc is benchmarked on the K-Radar dataset and demonstrates state-of-the-art performance in radar-based 3D occupancy prediction, outperforming LiDAR and camera-based methods. The method also shows superior performance in adverse weather conditions. The contributions of this work include the introduction of RadarOcc for 4D radar-based 3D occupancy prediction, the development of a novel pipeline to handle 4DRT challenges, and extensive experiments validating the method's performance. The paper also discusses the limitations of the current approach, including the lack of temporal information modeling and the limitation to two general semantics. The results show that RadarOcc provides robust 3D occupancy prediction under various conditions, making it a promising alternative to LiDAR for autonomous driving.