自动驾驶论文速递 | 端到端、Diffusion、VLM、OCC等方向~

Core Insights - The article discusses advancements in autonomous driving technology, focusing on knowledge-driven diffusion strategies and multimodal frameworks for improved performance in various driving scenarios [3][5][16][19][27][29][37][38]. Group 1: Knowledge-Driven Diffusion Policy (KDP-AD) - A new knowledge-driven diffusion policy (KDP) has been proposed, achieving success rates of 100%, 94%, and 90% in ramp merging, intersections, and roundabouts respectively, outperforming traditional methods [3][5]. - The KDP framework utilizes a mixture of experts (MoE) approach, allowing for modular and combinatorial strategy learning, enhancing knowledge reuse across different driving scenarios [5]. - Empirical validation shows that KDP significantly outperforms baseline methods in success rates, collision risk control, and generalization capabilities [5][9]. Group 2: SliceSemOcc Framework - The SliceSemOcc framework introduces a vertical slice-based multimodal 3D semantic occupancy prediction, improving mean Intersection over Union (mIoU) from 24.7% to 28.2%, a relative increase of 14.2% [16][19]. - It employs a dual-scale vertical slicing strategy to enhance feature extraction for both global and local contexts, addressing the representation of small objects effectively [19]. - Experimental results demonstrate significant improvements in mIoU across various datasets, particularly for small object categories [19][22]. Group 3: LatticeWorld Framework - The LatticeWorld framework integrates a lightweight multimodal large language model with Unreal Engine 5, achieving over 90 times efficiency improvement in industrial-grade scene generation [27][29]. - It supports multimodal inputs and dynamic multi-agent interactions, filling gaps in existing methods regarding interactivity and industrial efficiency [29]. - Performance validation indicates superior layout generation accuracy and visual fidelity compared to existing models, maintaining high creative quality [29][35]. Group 4: Ego3D-Bench and Ego3D-VLM - The Ego3D-Bench benchmark is introduced for ego-centric multi-view 3D spatial reasoning, enhancing visual language models' performance in various tasks [37][38]. - The Ego3D-VLM post-training framework improves 3D spatial reasoning capabilities, achieving an average accuracy increase of 12% in multiple-choice questions and a 56% reduction in RMSE for absolute distance estimation [38][41]. - The framework demonstrates adaptability across various state-of-the-art visual language models, confirming its generalization capabilities [38][41].