欢迎来到AI机器人的奇妙世界 — Welcome to the World of AI Robots

第 1 章：欢迎来到 AI 机器人的奇妙世界

Chapter 1: Welcome to the World of AI Robots

---

> 开篇场景 (Opening Scene) > 周末的早晨，8 岁的小诺跟着妈妈来到城市科技馆。展厅里，一个圆滚滚的扫地机器人正贴着墙根转圈；玻璃柜台上，一辆迷你自动驾驶汽车沿着磁条轨道缓缓前进；咖啡角里，六轴机械臂正把一只白瓷杯递到客人手中……小诺瞪大眼睛："这些机器人和手机里的 ChatGPT 有什么不一样？它们是怎么'看见'、'思考'、'动手'的？"妈妈笑着说："这就是我们今天要弄明白的事。"

> *On a weekend morning, eight-year-old Xiaonuo follows his mother to the city science museum. In the exhibition hall, a round robot vacuum cleaner is rolling along the wall base; on the glass counter, a miniature autonomous car is slowly moving along a magnetic strip track; at the coffee corner, a six-axis robotic arm is handing a white porcelain cup to a customer... Xiaonuo's eyes widen: "How are these robots different from ChatGPT on the phone? How do they 'see', 'think', and 'act'?" His mother smiles: "That's exactly what we're going to figure out today."*

---

1.1 三个核心问题：为什么要给 AI 一个"身体"？

1.1 Three Core Questions: Why Give AI a "Body"?

在开始动手之前，我们必须先回答三个问题。这三个问题会贯穿整本书，请把它们记在你的小本子上：

Before we start, we must answer three questions. These three questions will run through the entire book, so please write them down in your notebook:

问题一：什么是"具身智能"（Embodied Intelligence）？ *Question 1: What is "Embodied Intelligence"?*

> 专业定义 (Professional Definition)：具身智能是指智能体（agent）通过物理身体与真实环境进行实时交互，并在交互中学习、决策、适应的能力。简单说，就是"有身体的 AI"。

> *Embodied Intelligence refers to the ability of an agent to interact with the real environment in real-time through a physical body, and to learn, make decisions, and adapt during the interaction. Simply put, it's "AI with a body".*

问题二：为什么具身智能和 ChatGPT 这样的"聊天 AI"不一样？ *Question 2: Why is embodied intelligence different from "chat AI" like ChatGPT?*

| 维度 (Dimension) | 聊天 AI (Chat AI) | 具身 AI (Embodied AI) | |---|---|---| | 输入 (Input) | 文字、语音、图像 | 传感器流：触觉、力、距离、图像 | | 输出 (Output) | 文字、语音 | 电机动作、机械位移 | | 时间尺度 (Time scale) | 几百毫秒～几秒 | 毫秒级实时控制（1 ms = 0.001 s） | | 错误代价 (Error cost) | 答错可重说 | 撞墙、夹手、摔杯子 | | 核心能力 (Core ability) | 语言理解与生成 | 感知-决策-动作闭环 |

> 判断依据 (Judgment Basis)：两者的根本差异在于是否接触物理世界。ChatGPT 的世界是"文字构成的符号世界"，而具身 AI 的世界是"原子构成的真实世界"。在真实世界里，1 毫秒的延迟可能让机器人撞上一面墙，而一个错别字最多让人皱皱眉。

> *The fundamental difference lies in whether they touch the physical world. ChatGPT's world is a "world of symbols made of words", while embodied AI's world is a "real world made of atoms". In the real world, a 1-millisecond delay could make a robot crash into a wall, while a typo at most makes someone frown.*

问题三：什么是"世界模型"（World Model），它为什么重要？ *Question 3: What is a "World Model" and why is it important?*

世界模型是机器人大脑里的"内部沙盘"（Internal Sandbox）。它是一个可计算的程序，能根据当前的观察，预测未来几秒内世界会变成什么样。

A world model is an "internal sandbox" inside the robot's brain. It is a computable program that, based on current observations, can predict what the world will look like in the next few seconds.

> 举个例子 (Example)：扫地机器人面前有一张椅子。世界模型会告诉它："如果我继续向前走，0.5 秒后我会撞上椅腿；如果我向右转 30°，0.5 秒后我会安全通过。"——这就是预测。没有世界模型，机器人就只能在"撞-退-再撞"的循环里傻乎乎地摸索。

> *Example: A robot vacuum faces a chair. The world model tells it: "If I keep moving forward, in 0.5 seconds I'll hit the chair leg; if I turn 30° to the right, in 0.5 seconds I'll pass safely." This is prediction. Without a world model, the robot would foolishly grope around in a "crash-retreat-crash" loop.*

---

1.2 具身智能的三大要素：身体、大脑、环境

1.2 The Three Elements of Embodied Intelligence: Body, Brain, Environment

具身智能的三位一体闭环可以用一句话概括：

The trinity closed loop of embodied intelligence can be summarized in one sentence:

> 身体感知环境 → 大脑预测未来 → 身体执行动作 → 环境发生改变 → 身体再次感知…… > *Body senses environment → Brain predicts the future → Body executes action → Environment changes → Body senses again...*

这是一个永不停歇的反馈环（Feedback Loop），频率通常是 10 Hz～1000 Hz（即每秒 10 到 1000 次循环）。

This is a never-stopping feedback loop, usually running at 10 Hz to 1000 Hz (i.e., 10 to 1000 cycles per second).

1.2.1 身体（Body）：和世界"握手"的硬件

1.2.1 Body: The Hardware that "Shakes Hands" with the World

身体是机器人的物理外壳与机械结构。它由三类硬件组成：

The body is the robot's physical shell and mechanical structure. It consists of three types of hardware:

• **传感器（Sensors）**：相当于人的"五官"，把光、声、力、温度变成电信号。常见的有摄像头、麦克风、激光雷达（LiDAR）、陀螺仪、力矩传感器。

*Sensors: Equivalent to human "five senses", converting light, sound, force, and temperature into electrical signals. Common ones include cameras, microphones, LiDAR, gyroscopes, and torque sensors.*

• **计算单元（Computing Unit）**：相当于"大脑皮层"，通常是 CPU + GPU + NPU（神经网络处理器）的组合。嵌入式场景常用 NVIDIA Jetson、树莓派（Raspberry Pi）、Arduino 等。

*Computing Unit: Equivalent to the "cerebral cortex", usually a combination of CPU + GPU + NPU (Neural Processing Unit). Embedded scenarios commonly use NVIDIA Jetson, Raspberry Pi, Arduino, etc.*

• **执行器（Actuators）**：相当于人的"四肢和肌肉"，把电信号变成物理运动。常见的有舵机（Servo）、直流电机（DC Motor）、步进电机、气动肌肉。

*Actuators: Equivalent to human "limbs and muscles", converting electrical signals into physical movement. Common ones include servos, DC motors, stepper motors, and pneumatic muscles.*

> 小知识（Fun Fact）：人类的手有 23 个自由度（Degrees of Freedom, DoF），可以做出上百种抓取姿势。特斯拉 Optimus 机器人的单只手也有 11 个自由度，目标是复刻人类灵巧操作的 80%。 > > *Fun Fact: A human hand has 23 degrees of freedom (DoF) and can make hundreds of grasping gestures. The Tesla Optimus robot's single hand also has 11 degrees of freedom, aiming to replicate 80% of human dexterous operations.*

1.2.2 大脑（Brain）：世界模型 + 决策网络

1.2.2 Brain: World Model + Decision Network

机器人的大脑不是单一程序，而是一个模块化系统。最核心的两个模块是：

A robot's brain is not a single program, but a modular system. The two core modules are:

1. 世界模型（World Model, WM）：负责"想象未来"。它接收当前观察 $o_t$（observation at time $t$），输出未来若干步的预测 $\hat{o}_{t+1}, \hat{o}_{t+2}, \ldots$。数学上可以表示为： *World Model (WM): Responsible for "imagining the future". It receives the current observation $o_t$ and outputs predictions for several future steps $\hat{o}_{t+1}, \hat{o}_{t+2}, \ldots$ Mathematically, this can be expressed as:*

$$\hat{o}_{t+1}, r_{t+1}, d_{t+1} = f_\theta(o_t, a_t)$$

其中 $a_t$ 是当前动作，$r_{t+1}$ 是预测奖励，$d_{t+1}$ 是预测终止标志（如"撞墙了"）。 *Where $a_t$ is the current action, $r_{t+1}$ is the predicted reward, and $d_{t+1}$ is the predicted termination flag (e.g., "crashed into wall").*

2. 策略网络（Policy Network, $\pi$）：负责"选最好的动作"。它读取当前观察 $o_t$ 和世界模型的预测，输出最优动作 $a_t$。强化学习中常用 $\pi_\theta(a_t | o_t)$ 表示。 *Policy Network ($\pi$): Responsible for "choosing the best action". It reads the current observation $o_t$ and the world model's predictions, and outputs the optimal action $a_t$. In reinforcement learning, this is often denoted as $\pi_\theta(a_t | o_t)$.*

> 专业提示（Pro Tip）：现代具身 AI 系统（如 Google DeepMind 的 RT-2、NVIDIA 的 VIMA）通常把世界模型和策略网络联合训练（Joint Training），让它们相互促进——更准的预测带来更好的决策，更好的决策带来更准的预测。 > > *Pro Tip: Modern embodied AI systems (such as Google DeepMind's RT-2, NVIDIA's VIMA) usually jointly train the world model and the policy network, letting them promote each other—better predictions lead to better decisions, and better decisions lead to more accurate predictions.*

1.2.3 环境（Environment）：机器人必须遵守的"物理法则"

1.2.3 Environment: The "Laws of Physics" the Robot Must Obey

环境是机器人身处的真实物理世界。它有几个无法回避的特征：

The environment is the real physical world where the robot resides. It has several unavoidable characteristics:

• **连续性（Continuous）**：位置、速度都是实数（$\mathbb{R}^n$），不像棋盘游戏那样是离散的格子。

*Continuous: Positions and velocities are real numbers ($\mathbb{R}^n$), unlike board games that use discrete grids.*

• **高维性（High-dimensional）**：一张 1920×1080 的图像就有 200 多万像素；一秒钟的激光雷达点云可能包含 10 万个点。

*High-dimensional: A 1920×1080 image has over 2 million pixels; one second of LiDAR point cloud may contain 100,000 points.*

• **部分可观察（Partially Observable）**：机器人看到的永远只是世界的"一个切面"，墙后面的东西、桌角后面的杯子，它都看不见。

*Partially Observable: What the robot sees is always just a "slice" of the world; it cannot see things behind walls or cups behind table corners.*

• **动态性（Dynamic）**：人走来走去、光线变化、地面湿滑——世界每时每刻都在变。