https://auto-driving-blog.pages.dev/tags/sparseoccvla/ Recent content in SparseOccVLA on Elon's AD Insight https://auto-driving-blog.pages.dev/images/share.png https://auto-driving-blog.pages.dev/images/share.png Hugo zh-cn Mon, 27 Jul 2026 00:00:00 +0000 https://auto-driving-blog.pages.dev/posts/paper-reading/sparseoccvla%E7%B2%BE%E8%AF%BB/ Mon, 27 Jul 2026 00:00:00 +0000 https://auto-driving-blog.pages.dev/posts/paper-reading/sparseoccvla%E7%B2%BE%E8%AF%BB/ SparseOccVLA 用稀疏占用查询作为视觉与语言之间的单一桥梁，彻底摆脱 ViT patch token，让 LLM 在统一框架内同时完成场景理解、占用预测和轨迹规划。在 OmniDrive-nuScenes 上 CIDEr 提升 7%，Occ3D-nuScenes mIoU 提升 0.5，nuScenes 规划指标达 SOTA。