你好,世界!——为什么我们要关心具身智能与AI的世界模型
第 1 章 你好,世界!——为什么我们要关心具身智能与 AI 的世界模型
Imagine you ask a super-smart AI to "make me a cup of coffee." It can write you a beautiful essay about coffee brewing, design an infographic of latte art, even explain the chemistry of extraction. But the moment you put a real mug on a real table and ask it to actually *do* it, something is missing. This chapter starts from that gap — and from there we'll discover why the next decade of AI will be defined not by bigger thinking machines, but by machines that *live in the world*.
---
1.1 一杯咖啡引出的问题:纯思考 AI 为何"动不了手"
Let's start with a simple experiment. In your mind, picture a coffee cup. Now picture yourself reaching out, wrapping your fingers around the handle, lifting it up, tilting it slightly so the liquid doesn't spill. Easy, right? You didn't even have to think hard.
Now ask a powerful Large Language Model (LLM) — like the kind that writes essays, code, and poems — to do the same task in the *real* world. Most of them have no arms, no eyes, and no sense of touch. Even if you connected them to a robot, they often struggle with tasks that any 5-year-old can do in seconds. A famous 2024 Stanford robotics study showed that state-of-the-art models could plan dialogues about pouring water, but when attached to a real robot arm, they knocked over the cup more than 40% of the time.
Why? Because writing a paragraph is not the same as manipulating a physical object. The paragraph only requires *symbols*; the pouring requires a body, a feel for physics, and an internal simulation of what will happen next.
This book argues that the leap from "AI that thinks about the world" to "AI that lives in the world" is as significant as the leap from calculators to personal computers. We will call this leap Embodied AI / 具身智能 (jùshēn zhìnéng) — and at its heart lies a magical capability called a World Model / 世界模型 (shìjiè móxíng).
> 💡 关键术语 / Key Term > 具身智能 (Embodied AI): AI 系统的智能来源于其与物理世界的实时交互 —— 它的"身体"和"大脑"是同一个系统不可分割的两部分。 > 世界模型 (World Model): 大脑(生物或人工)对"世界如何运作"建立的内部模拟器,能预测"如果我这样做,会发生什么"。
Three real-world examples will make this concrete:
| 场景 / Scene | "纯思考" AI 表现 / Pure-Thinking AI | 为什么失败 / Why It Fails | |---|---|---| | 自动驾驶识别"罕见卡车" / Self-driving car sees rare truck | 大模型可能从未在训练数据中见过这种卡车与骑自行车的行人同时出现 | 缺乏对"如果我现在变道"的物理直觉 | | 家务机器人整理餐桌 / Home robot clears dinner table | 可以生成一份完美的清单,但抓杯子时容易捏碎 | 没有"杯子是空的还是满的、是瓷还是玻璃"的触觉/重量模型 | | AI 替你看病 / AI diagnoses rare disease | 文字描述上非常专业,但无法拍胸口听声音 | 缺乏多模态感知 + 在闭环中迭代决策的能力 |
这些案例并不是说 LLM 不聪明 —— 它极其聪明。问题在于:the kind of intelligence you need to *describe* a task and the kind you need to *perform* it are fundamentally different.
1.1.1 从"语言游戏"到"物理游戏"
哲学家把这个分野叫做 the symbol grounding problem / 符号接地问题: 词语只有在它们指向真实世界的物体、动作、力与感觉时,才是"有意义的"。一个只读过字典的人和一个摸过滚烫杯子的人,对"烫"这个字的表征完全不同。
今天的 LLM 在某种意义上是"博览群书但从未下过雨"的学者。它对世界的理解来自互联网上的文字,而非来自 *interacting with* the world. This is why robot labs at Stanford, Google DeepMind, and Tesla have started pouring billions of dollars into what they now call embodied foundation models / 具身基础模型 — models designed from the start to handle perception, prediction, and action.
1.1.2 一句话总结这节内容
> The bottleneck for AI is no longer raw thinking. It is *action grounded in a body that understands the world*.
---
1.2 你已经拥有世界上最强的"世界模型"——你的大脑
Before we look at what AI's world models should look like, let's appreciate the one you already carry in your head. Every morning when you brush your teeth, your brain runs a tiny version of the simulation we are about to discuss.
1.2.1 刷牙的"内心独白"
Pick up your toothbrush. You don't have to look at the mirror to know where your mouth is. You don't have to look at the tube to know how hard to squeeze. You don't have to do math to know that if you tip the brush down, the toothpaste will fall on the floor. This is your 世界模型 at work:
- **感知 (Perception)**: 牙刷的重量、牙膏管的软硬、镜子里嘴唇的开合。
- **表征 (Representation)**: 在你心里"长出"一张嘴巴的三维地图,里面有牙齿、舌头的位置。
- **预测 (Prediction)**: 想象刷头向左移动 1 厘米,会碰到臼齿;向右偏 0.5 厘米,会戳到牙龈。
- **决策 (Decision)**: 选择路径 A —— 由外向内、上下刷。
- **行动 (Action)**: 小肌肉执行决策。
- **反馈 (Feedback)**: 牙龈出血,意味着下次力度调小。
整个链条在你"无意识"的状态下完成,但它的确涉及一个内部模拟器。如果明天你换了一把电动牙刷,你的世界模型会在两三次刷动后自动 *更新*——它已经预测到这个新工具的反作用力会比旧的不同。这就是世界模型的核心:a learnable, updatable simulator of "how the world responds to my actions".
1.2.2 大脑为什么这么厉害
Cognitive scientists often call the brain's simulator a generative model / 生成式模型: 它接收当前状态(看到的、摸到的)和候选动作(我想做的),输出对未来的预测(接下来会发生什么)。三大神经科学原理值得记住:
1. 预测编码 (Predictive Coding): 大脑不停地做预测,并把预测错误当成学习信号。瑞士苏黎世大学的神经科学家对这个理论有长达 30 年的研究。 2. 内模仿 (Internal Rehearsal): 运动员、钢琴家在真正的动作之前,已经在心里"预演"了数百遍。 3. 反事实推理 (Counterfactual Reasoning): 我虽然没有真的把杯子推下桌,但我能想象它摔碎的样子——这个"想象"改变了我对现实的选择。
This is precisely what we want AI's world models to do. Whether a model is implemented by a Transformer, a diffusion network, or a spiking neural network, it should accept a *state* and an *action*, and output a *predicted next state* — plus an estimate of how confident it is.
1.2.3 一句话总结这节内容
> 你的大脑是一个永不停歇的世界模型:感知 → 表征 → 预测 → 决策 → 行动 → 反馈。
---
1.3 什么是世界模型?——用乐高积木讲清楚
Let's move from biology to engineering. The next era of AI won't be built only from chat interfaces. It will be built from world models that AI can run inside its silicon brain to "try things out" before committing real actions.
1.3.1 三个层级定义
为了让你不会在术语里迷路,我们用三层"乐高积木"来定义世界模型:
Layer 1 — 物理动力学模型 (Physics Dynamics Model) "如果不松手,球会停;如果倾斜超过 30°,水会洒。" 这是最古老的一类世界模型。60 年代的控制系统就是用它预测汽车、火箭的姿态。代表工具:物理仿真器 MuJoCo、Isaac Sim、Bullet。
Layer 2 — 感知驱动的世界模型 (Perception-Grounded World Model) 直接来自视频与传感器的学习。给定一段自动驾驶视频,模型学到"红灯后车会减速"这种隐含规则。代表研究:Google DeepMind 的 V-JEPA、Waabi 的 Waymax、Wayve 的 GAIA。
Layer 3 — 语义-动作耦合的世界模型 (Semantically-Coupled World Model) 不仅知道物理,还理解意义:这一团像素是"护士",那个动作是"在抢救",下一步最可能是"推急救车"。代表方向:具身大模型 PaLM-E、RT-2、π₀ by Physical Intelligence。
一个有用的比喻: 把世界模型想成游戏中的 *Predictive Engine / 预测引擎*。
在《我的世界》(Minecraft) 这类游戏里,运行模型在每一帧会预测"如果我现在点鼠标右键敲这块石头,下一帧石头会变成什么"。真正的世界模型,就是让 AI 像玩家一样*通过行动探索和理解*世界,并且让模型在探索中不断被改写。
1.3.2 一张图看懂工作流
下面这张图描述了世界模型每时每刻在后台运行的内部循环。后面章节会更细致地拆解,今天只需要把图像记住:
```mermaid flowchart LR A[State s_t<br/>当前状态] --> C[World Model<br/>世界模型] B[Action a_t<br/>候选动作] --> C C --> D[Predicted s_{t+1}<br/>下一帧预测] C --> E[Reward / Surprise<br/>奖励或惊讶度] D --> F[Decision Module<br/>决策模块] E --> F F --> B F --> G[Real Action<br/>真实执行] G --> A ```
注意箭头 *G → A*:真实动作改变了真实世界的状态,然后新的状态又流回世界模型。这是闭环 (closed loop)的关键 —— AI 不是在沙盒里空想,而是活在"模型 ↔ 真实世界"的双向管道里。
1.3.3 一个可以亲手做的思维实验
下面这个清单不是代码,而是给你未来工程师的"建模直觉训练":
1. 选一件你每天都做的事(例如穿鞋)。 2. 写出 5 个"隐含预测": *If I bend my knee 90° and lift my heel, the back of the shoe will open enough for my foot to slide in.* 3. 假设你从没穿过鞋,写出你*会犯错*的 3 个反事实: *If I try to put my foot in from the front, my heel will be crushed against the back of the shoe.* 4. 现在想象 — 如果你要 *teach a robot* to put on a shoe, what state information would the robot need? 高度?脚的角度?摩擦力?
这个练习就是你第一次"以工程师身份思考世界模型"。我们将在第 6 章回顾它,并将其形式化成一个 state-action-next-state tuple / 状态-动作-下一状态三元组。
1.3.4 一句话总结这节内容
> 世界模型就是一个 *learnable simulator of consequences*——一个能被数据驱动并持续更新的"后果预测器"。
---
1.4 什么是具身智能?——从波士顿动力到 Figure 02
Now we ascend one level. A world model is the *mind* of the intelligent agent. Embodied AI / 具身智能 is the *whole package* — mind plus body plus the wiring between them.
1.4.1 一个现代工作定义
> 具身智能 / Embodied AI 是指 AI 系统的智能行为来源于其 身体(形态、传感器、执行器) 与 环境 持续、实时的交互过程,并通过内部世界模型对外进行预测、规划与反思。
我们把这个定义拆成三个支柱:
- **形态 (Morphology)**: 机器人的形状、关节数量、材质 —— 会从根本上改变它能做什么。一个没有手指的假手机器人很难"倒咖啡",但可以"按按钮"。
- **感知 (Perception)**: 多模态传感器 —— 摄像头、LiDAR、麦克风、力矩传感器、触觉皮肤、肌电信号 —— 共同构建对环境的实时表征。
- **行为 + 学习 (Behavior + Learning)**: 既包括经典控制(PID、运动规划),也包括通过与世界模型的互动进行强化学习、模仿学习、自进化。
1.4.2 产业现状:四条前线
下面这张产业图谱帮你一眼看清具身智能最有活力的几个方向:
| 方向 / Direction | 代表玩家 / Players | 当前瓶颈 / Bottleneck | |---|---|---| | 自动驾驶 / Autonomous Driving | Waymo, Tesla FSD, Wayve, 小马智行, Apollo | 长尾场景与"陌生拓扑" | | 工业机器人 / Industria