跳到主要内容Skip to main content

世界模型World Model

World Model

  • 模型能力Model capabilities
  • 第二站 · 上手Station 2 · Hands On
  • 引用核验 2026-11-03refs checked 2026-11-03
解决什么问题What problem it solves

让模型在内部预测「世界接下来会怎样」,而不只是描述当前画面。Let a model predict what the world will do next, not merely describe the current frame.

这是什么What it is

世界模型是对环境状态与变化的压缩表示:给定当前状态与动作,预测下一状态、画面或物理结果。

视频生成、3D 生成与机器人规划都在用这个概念,但各自「世界」的粒度不同。

A world model is a compressed representation of environment state and change: given the current state and an action, it predicts the next state, frame or physical outcome.

Video generation, 3D generation and robot planning all use the term, but at different granularity.

为什么需要知道Why it matters

它把「生成一段视频」和「理解物理规律」区分开:画面好看不代表模型真的知道物体会怎么动。It separates "generating a video" from "understanding physics": good-looking frames do not prove the model knows how objects move.

出处Sources
  • Ha & Schmidhuber · World Models(arXiv 1803.10122)Ha & Schmidhuber, World Models ↗
  • DeepMind · Genie 2 官方介绍DeepMind · Genie 2 ↗