Selected work

Representative Papers

PrismAD: Decoupled Planning via Semantic Mixture-of-Planners for End-to-End Autonomous Driving

Jul 11, 2026

This work addresses the limitations of existing end-to-end autonomous driving planners, which couple heterogeneous scene information into a unified representation space, thereby hindering specialized modeling of critical factors such as agent interactions, road geometry, and driving intent. To overcome this, the authors propose PrismAD, a novel semantic-driven decoupled planning framework. PrismAD introduces a semantic mixture planner that decomposes inputs into three distinct semantic groups—interaction, geometry, and intent—each processed by dedicated expert networks with disjoint parameters. A semantic-aware router then adaptively fuses these outputs using sparse Top-K activation combined with a noisy gating mechanism. This design enhances routing robustness while reducing computational overhead, achieving competitive performance on both the nuScenes open-loop dataset and the NeuroNCAP closed-loop benchmark.

0 citationsRead paper

RadarXFormer: Robust Object Detection via Cross-Dimension Fusion of 4D Radar Spectra and Images for Autonomous Driving

Mar 16, 2026

This work proposes a 3D object detection framework that fuses raw 4D millimeter-wave radar spectrograms with RGB images to address the performance degradation of camera and LiDAR perception under adverse weather and lighting conditions. The method innovatively leverages raw radar spectrograms directly to construct a compact representation that preserves complete 3D geometric information. A cross-modal Transformer mechanism is designed to integrate multi-scale 3D spherical features from radar with 2D image features, enhancing spatial consistency while reducing data redundancy. Evaluated on the K-Radar dataset, the approach significantly improves detection accuracy and robustness in complex environments without compromising real-time inference capability.

0 citationsRead paper

GraphBrep: Learning B-Rep in Graph Structure for Efficient CAD Generation

Jul 07, 2025

To address geometric-topological misalignment in B-Rep generation—leading to redundant representations and high computational overhead—this paper proposes the first end-to-end B-Rep generation method based on undirected weighted graphs. It explicitly models B-Rep surface topology as a graph structure and jointly learns the geometric-topological joint distribution via graph neural networks and graph diffusion models, enabling decoupled geometric and topological modeling within a compact representation. By eliminating redundant intermediate representations inherent in conventional pipeline-based approaches, the framework significantly reduces computational cost. Extensive evaluation on multiple large-scale CAD datasets demonstrates that our method achieves state-of-the-art generation accuracy while reducing training and inference time by up to 31.3% and 56.3%, respectively—marking the first approach to substantially improve B-Rep generation efficiency without compromising fidelity.

0 citationsRead paper

Recent publications

Latest Papers

PrismAD: Decoupled Planning via Semantic Mixture-of-Planners for End-to-End Autonomous Driving

Jul 11, 2026

This work addresses the limitations of existing end-to-end autonomous driving planners, which couple heterogeneous scene information into a unified representation space, thereby hindering specialized modeling of critical factors such as agent interactions, road geometry, and driving intent. To overcome this, the authors propose PrismAD, a novel semantic-driven decoupled planning framework. PrismAD introduces a semantic mixture planner that decomposes inputs into three distinct semantic groups—interaction, geometry, and intent—each processed by dedicated expert networks with disjoint parameters. A semantic-aware router then adaptively fuses these outputs using sparse Top-K activation combined with a noisy gating mechanism. This design enhances routing robustness while reducing computational overhead, achieving competitive performance on both the nuScenes open-loop dataset and the NeuroNCAP closed-loop benchmark.

0 citationsRead paper

RadarXFormer: Robust Object Detection via Cross-Dimension Fusion of 4D Radar Spectra and Images for Autonomous Driving

Mar 16, 2026

This work proposes a 3D object detection framework that fuses raw 4D millimeter-wave radar spectrograms with RGB images to address the performance degradation of camera and LiDAR perception under adverse weather and lighting conditions. The method innovatively leverages raw radar spectrograms directly to construct a compact representation that preserves complete 3D geometric information. A cross-modal Transformer mechanism is designed to integrate multi-scale 3D spherical features from radar with 2D image features, enhancing spatial consistency while reducing data redundancy. Evaluated on the K-Radar dataset, the approach significantly improves detection accuracy and robustness in complex environments without compromising real-time inference capability.

0 citationsRead paper

GraphBrep: Learning B-Rep in Graph Structure for Efficient CAD Generation

Jul 07, 2025

To address geometric-topological misalignment in B-Rep generation—leading to redundant representations and high computational overhead—this paper proposes the first end-to-end B-Rep generation method based on undirected weighted graphs. It explicitly models B-Rep surface topology as a graph structure and jointly learns the geometric-topological joint distribution via graph neural networks and graph diffusion models, enabling decoupled geometric and topological modeling within a compact representation. By eliminating redundant intermediate representations inherent in conventional pipeline-based approaches, the framework significantly reduces computational cost. Extensive evaluation on multiple large-scale CAD datasets demonstrates that our method achieves state-of-the-art generation accuracy while reducing training and inference time by up to 31.3% and 56.3%, respectively—marking the first approach to substantially improve B-Rep generation efficiency without compromising fidelity.

0 citationsRead paper