OVOW：从单目视频重建可仿真 4D 物理场景
一段视频即可生成可交互的物理仿真场景，为机器人提供低成本训练数据。
机器人训练需要海量真实环境数据，但人工建模成本高。OVOW 系统能从一段普通单目视频中，自动拆出独立物体、补全几何、恢复尺寸和运动，并按重力与支撑关系组装，生成可直接进入物理引擎的 4D 网格世界。该研究由清华、中科大等团队完成，已被 ECCV 2026 接收。
正文摘录
.jpg)  Once robots leave the lab, training data is no longer just a question of "how much": whether the environment can cover the scale, occlusion, contact, and motion of the real world directly determines whether the policy can be deployed. This is why Real2Sim is receiving renewed attention — if ordinary videos can be turned into interactive simulation scenes, the cost of data collection would be far lower than manual modeling or specialized equipment. The challenge is that the world seen in a video is not…