机器人学习：方法与经典工作Robot Learning: Methods & Classic Works
四章、10 节，把 Part V 的工具真正用到机器人上：模仿学习与数据讲怎样从示教里学策略，从行为克隆和 DAgger 走到 ACT 与 Diffusion Policy，再回到示教数据本身从哪里来；机器人强化学习讲怎样在仿真里大规模训练、再跨过现实差距搬上真机；Model-based 方法讲怎样先学一个动力学模型，再在模型里规划或训练策略，从 PILCO、PETS 走到 Dreamer、TD-MPC 和学习型 MPC；经典工作用一条可交互的时间线，把这些方法放回它们出现的年代。每一节都有在浏览器里现场运行的实验，大多是现场训练的策略和模型。 Four chapters and 10 lessons that put the tools of Part V to work on real robots: imitation learning and data covers learning a policy from demonstrations, from behaviour cloning and DAgger to ACT and Diffusion Policy, and back to where the demonstrations come from; robot reinforcement learning covers training at scale in simulation and crossing the reality gap to the real machine; model-based methods cover learning a dynamics model first and then planning or training a policy inside it, from PILCO and PETS to Dreamer, TD-MPC and learned-model MPC; landmark works puts these methods back where they appeared, on an interactive timeline. Every lesson has labs that run live in your browser, most of them policies and models trained on the spot.
1四章之间的关系How the four chapters connect
Part VI 主要建立在 Part V 之上。模仿学习（22 章）用到生成模型（19 章）的 VAE、扩散与 flow matching 和架构（17 章）的 Transformer，遥操作一节还用到 Part I–II 的雅可比、SE(3) 与逆运动学；机器人强化学习（23 章）建立在 PPO/SAC（20 章）和不确定性（21 章）之上，教师–学生蒸馏还借用了 22.1 节的 DAgger；Model-based 方法（24 章）与前两章互不依赖，它用到 RNN、VAE、PPO/SAC 与不确定性，以及 Part I 的采样优化和 Part IV 的状态空间与 MPC。第 25 章的时间线回顾的是整个 Part V 与 Part VI，图里只画出它到第 22–24 章的边。把鼠标放在（或点击）任意一个方块上：琥珀色是它依赖的部分，紫色是用到它的部分；虚线框是前面的章节或后续部分，本部分各章之间的依赖画成右侧的弧线。每一节具体依赖哪几节，见第 3 部分。
Part VI rests mainly on Part V. Imitation learning (chapter 22) uses the VAEs, diffusion and flow matching of generative models (chapter 19) and the transformers of architectures (chapter 17), and its teleoperation lesson also uses the Jacobian, SE(3) and inverse kinematics from Parts I–II; robot reinforcement learning (chapter 23) builds on PPO/SAC (chapter 20) and uncertainty (chapter 21), and its teacher–student distillation borrows DAgger from lesson 22.1; model-based methods (chapter 24) do not depend on the two chapters before them, and use RNNs, VAEs, PPO/SAC and uncertainty, plus sampling-based optimization from Part I and the state space and MPC from Part IV. The timeline of chapter 25 looks back over all of Parts V and VI; the graph draws only its edges from chapters 22–24. Hover (or tap) any box: amber shows what it builds on, purple shows what builds on it; dashed boxes are earlier chapters or later parts, and dependencies between this part's chapters are drawn as arcs on the right. Section 3 lists exactly which lessons each lesson builds on.
2各章一览The chapters
★ 标记的是核心节，点击标题即可进入（灰色的是尚未上线的节）。每章都会写出它的"一句话直觉"和在机器人中的主要用途。
★ marks core lessons; click a title to open it (greyed-out lessons are not live yet). Each chapter lists its one-line intuition and its main uses in robotics.
3逐节内容与前置Each lesson and what it builds on
每节一句话说明它讲什么，下面列出该节自己标注的前置节。青色边框的是本部分内的前置节，其余来自 Part I–V；点击即可跳转。
Each lesson gets one line on what it covers, followed by the prerequisites the lesson itself lists. Prerequisites with a cyan border are inside this part; the rest come from Parts I–V. Click one to jump there.
4按目标选路线Pick a route by goal
时间有限的话，可以按你的方向挑一条路线先走通，再回头补全。路线里跳过的节，遇到时看第 3 部分列出的前置即可。
Short on time? Pick the route that matches your direction, walk it end to end, then come back and fill in the rest. When a route skips a lesson you turn out to need, the prerequisites in section 3 show where to look.