[{"data":1,"prerenderedAt":51},["Reactive",2],{"news-story-news-1792368356004-vsofl":3},{"id":4,"title_zh":5,"title_en":6,"summary_zh":7,"summary_en":8,"takeaways_zh":9,"takeaways_en":14,"content_zh":19,"content_en":20,"external_links":21,"source":28,"category":35,"impact_score":36,"views_count":37,"likes_count":38,"related_entities":39,"tags":43,"image_url":50},"news-1792368356004-vsofl","同济与腾讯提出 3D 世界模型视角解耦表征 GeoState：分离相机畸变与场景演变，Waymo 跨帧预测零尺度漂移","Tongji University & Tencent Introduce GeoState: View-Decoupled 3D World Modeling Prevents Camera Scale Drift Across Waymo and KITTI Rollouts","现有 3D 世界模型多直接采用几何基础模型（GFM）特征作为状态，但视角依赖性导致相机运动与场景演变相互纠缠，长链路外推极易发生尺度漂移崩溃。同济大学与腾讯团队发表 GeoState（arXiv:2610.22343，项目主页上线）。该框架利用场景 Tokenizer 将多视角几何特征聚合为视角解耦的世界状态，自回归模型仅推演绝对场景状态，在 Waymo 自动驾驶长程深度预测中全程保持单一上下文尺度，且零样本泛化至 KITTI 与 nuScenes。","Modern 3D world models typically use view-dependent features from geometry foundation models (GFMs) as predictive states, entangling camera-induced variations with physical scene dynamics. Tongji University and Tencent introduced GeoState (arXiv:2610.22343). By aggregating multi-view GFM features into a view-decoupled 3D scene state, GeoState models physical scene evolution independently of camera viewpoints, maintaining a single context-derived scale throughout rollouts without per-frame scale realignment.",[10,11,12,13],"同济大学与腾讯团队提出 3D 世界模型视角解耦表征 GeoState（arXiv:2610.22343），根本性解决 GFM 特征视点依赖造成的长程尺度漂移","通过场景 Tokenizer 将环视几何特征聚合为视角解耦状态，自回归模型推演物理演化，几何解码器按需响应目标相机光线查询","在 Waymo 自动驾驶长程深度预测（t+0.2s 至 t+2.0s）中全程维持单一上下文尺度，全面超越 DINO-Foresight 与 VGGT-World","展现出强大的跨传感器泛化能力，零样本无缝迁移至 KITTI 与 nuScenes 复杂路况",[15,16,17,18],"Tongji University and Tencent introduced GeoState (arXiv:2610.22343) for view-decoupled 3D world modeling","Separates physical scene evolution from camera-dependent observations via a dedicated scene tokenizer and ray-based geometry decoder","Maintains a single context-derived scale throughout multi-second rollouts on Waymo without per-frame scale alignment, outperforming VGGT-World and FR3D","Demonstrates robust zero-shot cross-sensor generalization across KITTI and nuScenes datasets","### 核心背景与行业痛点\n在自动驾驶端到端规划与机器人 3D 具身仿真中，构建具备物理前瞻能力的 3D 世界模型是核心前沿方向。近期研究多尝试将几何基础模型（GFM，如 DINOv2、Depth Anything、DUSt3R）的隐层特征直接作为预测世界状态。然而，**这些 GFM 特征高度依赖特定的相机视角（View-Dependent）**——当自动驾驶车辆移动或变道时，相机视角的几何畸变与场景本身的动态变化严重耦合，导致自回归世界模型在多帧推演中迅速累积尺度误差，甚至需要每一帧进行人工强制尺度对齐（Per-frame Scale Alignment）才能勉强维持稳定。同济大学与腾讯团队在 **`arXiv:2610.22343`** 中给出了根本性的视角解耦方案 **`GeoState`**。\n\n### 架构亮点与底层机制\n1. **视角解耦场景 Tokenizer（Scene Tokenizer）**：\n   - 系统首先利用多视角 GFM 提取环视观测特征，随后通过跨视图聚合 Tokenizer 将其映射为一个**绝对视角解耦的 3D 场景全局状态（View-Decoupled Scene State）**，彻底剥离特定摄像机的光线投射参数；\n2. **单上下文尺度自回归状态推演与按需光线解码（Geometry Decoder）**：\n   - 自回归世界模型仅负责在解耦的状态隐空间中预测未来场景的自然物理演化，而相机位姿转移作为独立输入；\n   - 在最终渲染输出时，轻量级几何解码器依据指定的目标相机光线（任意虚拟视角或预测位姿）按需查询解耦状态，**在整个多秒 Rollout 过程中全程维持由初始上下文锚定的单一真实尺度**，彻底摒弃了逐帧尺度对齐。\n\n### 权威 Benchmark 与实测跑分对比\n- **Waymo 自动驾驶长程深度预测（Table 1, Long-Horizon Forecasting）**：\n  - 在 $t+0.2\text{ s}$ 至 $t+2.0\text{ s}$ 的连续长程深度预测中，`GeoState` 在不进行逐帧尺度对齐的前提下，深度绝对相对误差（AbsRel ↓）与阈值准度（δ1 ↑）全面击败 `DINO-Foresight`、`VGGT-World` 与 `FR3D`，在 $t+2.0\text{ s}$ 远期推演中保持卓越的几何轮廓稳定性；\n- **跨数据集零样本泛化能力（Table 2, KITTI & nuScenes Zero-Shot）**：\n  - 模型在 Waymo 上训练完成后，直接在未经微调的 **KITTI**（$t+0.6\text{ s}$ 至 $t+2.0\text{ s}$）与 **nuScenes**（$t+0.75\text{ s}$ 至 $t+2.5\text{ s}$）上进行零样本评测，其深度预测准确度显著超越专门在多传感器上做过对齐的对比模型。\n\n### 开发者实战落地与开箱指南\n- **项目主页与演示**：官方已开放项目演示主页（`https://hongchang-liu.github.io/GeoState-project-page/`）。\n- **具身世界模型构建启示**：构建端到端自动驾驶仿真器或双臂机器人物理模拟器的团队，不应将多视角视频帧直接塞入扩散网络；采用 GeoState 范式先抽离视点无关的几何状态，能够使长期物理预测误差降低数倍，并支持以任意虚拟摄像机自由视角漫游回放。","### 核心背景与行业痛点\nIn autonomous driving and robotics, 3D world models are essential for forward planning and counterfactual simulation. Recent approaches adopt latent features from geometry foundation models (GFMs) as world representations. However, because GFM features are inherently view-dependent, camera-induced transformations are entangled with true scene dynamics. This entanglement leads to severe scale drift and geometric instability during autoregressive rollouts, forcing previous models to rely on heuristic per-frame scale alignments. Researchers from Tongji University and Tencent introduced **`GeoState` (`arXiv:2610.22343`)** to decouple physical scene dynamics from sensor viewpoints.\n\n### 架构亮点与底层机制\n1. **View-Decoupled Scene Tokenizer**:\n   - Aggregates multi-view GFM embeddings into a global, viewpoint-independent 3D scene state representation, isolating scene geometry from individual camera projections;\n2. **Single Context Scale Autoregression & Ray-Based Decoding**:\n   - An autoregressive world model predicts future scene state transitions independently of camera movement. A lightweight geometry decoder subsequently queries the predicted state using arbitrary camera rays, preserving a single continuous scale across the entire rollout without per-frame realignment.\n\n### 权威 Benchmark 与实测跑分对比\n- **Waymo Long-Horizon Depth Forecasting (Table 1)**:\n  - Across forecasting horizons from $t+0.2\\text{ s}$ up to $t+2.0\\text{ s}$, `GeoState` consistently outperforms `DINO-Foresight`, `VGGT-World`, and `FR3D` on AbsRel and $\\delta_1$ accuracy metrics without needing artificial per-frame scale alignment;\n- **Zero-Shot Transfer Across KITTI and nuScenes (Table 2)**:\n  - Generalizes zero-shot to KITTI ($t+0.6\\text{ s}$ to $t+2.0\\text{ s}$) and nuScenes ($t+0.75\\text{ s}$ to $t+2.5\\text{ s}$), demonstrating superior cross-sensor physical generalization.\n\n### 开发者实战落地与开箱指南\n- **Project Resources**: Project demos are available at `https://hongchang-liu.github.io/GeoState-project-page/`.\n- **Engineering Takeaways**: When developing 3D simulation or driving world models, decoupling camera observation parameters from internal geometric state evolution prevents long-horizon drift and enables arbitrary novel-view rendering of future rollouts.",[22,25],{"title":23,"url":23,"type":24},"https://hongchang-liu.github.io/GeoState-project-page/","other",{"title":26,"url":26,"type":27},"https://arxiv.org/abs/2610.22343","arxiv",{"platform":29,"author_name":30,"author_handle":31,"author_avatar":32,"tweet_url":33,"published_at":34},"twitter","Tongji University & Tencent","@huggingface","https://unavatar.io/x/huggingface","","2026-10-19T00:05:56.004Z","Research",90,32,0,{"models":40,"agents":41,"vendors":42},[],[],[],[44,45,46,47,48,49],"GeoState","3D World Models","Autonomous Driving","Tencent","Waymo","GFM","https://arxiv.org/html/2610.22343v1/geostate_framework.svg",1792919796573]