aresbit 具身智能导论 — Knowledge Reference & Tutor
This skill answers questions about embodied intelligence by grounding them in
《具身智能导论》, a 17-chapter Chinese textbook (translated & expanded from UC
Berkeley CS294-291 "Robots That Learn", Spring 2026). The full chapters live in
references/book/chXX_*.md, with a 中英术语表 in references/book/glossary.md.
How to use this skill
The book is ~1 MB across 17 chapters — do NOT read it all. Each chapter is a self-contained ~50 KB unit. Work like this:
- Route the question to chapter(s) using the index below (match by topic, keyword, paper, or method name).
- Read only the relevant chapter file(s) with the Read tool. For a focused question, Grep the chapter for the specific term first, then read around it. For multi-topic questions, read the 2-3 chapters that apply.
- Answer grounded in the chapter: use its derivations, its notation, its
中英对照 terminology (cross-check
glossary.md), and cite where it comes from — e.g. "见第9章 §二 卡尔曼滤波" — so the user can go deeper. - Preserve math fidelity: reproduce the book's LaTeX notation exactly; don't silently re-derive with different symbols. If you extend beyond the book, say so explicitly.
- Code questions: each chapter has a §四 算法与代码实现 with PyTorch-flavored implementations — start from those rather than inventing a new API.
If a question falls outside the book's scope, say so plainly, then answer from general knowledge while flagging that it's not from the book. Don't pretend the book covers something it doesn't.
Chapter routing index
Match the user's question to the chapter(s) by topic / keyword / paper, then read
that file from references/book/.
| Chapter file | 主题 | Route here when the question is about… |
|---|---|---|
ch01_导论.md |
导论 | why robots should learn, 具身假设/embodiment, course overview, history of embodied AI, sense-plan-act vs learning |
ch02_生物运动力学.md |
生物运动力学 | CPG 中枢模式发生器, Matsuoka 振荡器, Kuramoto 同步, 步态/gait, 占空比, DMP 动态运动基元, locomotion biomechanics (Ramdya & Ijspeert) |
ch03_机器人机构学.md |
机构学 | 李群 SO(3)/SE(3), 旋转矩阵, 旋量/twist, 指数映射, 正/逆运动学, 雅可比, 拉格朗日 & 牛顿-欧拉动力学, DH 参数 |
ch04_扩散模型入门.md |
扩散模型 | DDPM, 前向/反向扩散, score matching, denoising, 流匹配 flow matching, normalizing flows, ELBO (Papamakarios, Lipman) |
ch05_人手与机器人手.md |
灵巧手 | 人手解剖学, 灵巧手设计, 欠驱动 underactuation, 协同 synergy, LEAP/软体手, 机械智能 (Piazza et al. 2019) |
ch06_本体感觉与触觉感知.md |
触觉感知 | 触觉传感器, GelSight, 视触觉, 皮肤力学, 本体感受 proprioception, 机械感受器 (Jones, Gardner) |
ch07_运动控制的发展视角.md |
发育运动控制 | 发育认知, 运动基元, CPG 视角, 跨模态视觉行走, 自监督运动 (Loquercio; Smith & Gasser) |
ch08_机器人动力学与控制.md |
动力学与控制 | LQR, MPC 模型预测控制, 轨迹优化, 阻抗/导纳控制, 运动规划, 最优控制 (Kawato, Flanagan) |
ch09_计算神经科学与预测控制.md |
预测控制 | 前向/逆向模型, 小脑模型, 卡尔曼滤波 Kalman filter, 贝叶斯估计, 精度加权, 预测控制, Land 茶实验 |
ch10_视频世界模型.md |
世界模型 | 视频预测, 潜在动态模型, world action models (WAM), Track2Act, 点轨迹, 零样本策略 (Ye; Bharadhwaj) |
ch11_强化学习.md |
强化学习 | MDP, 贝尔曼方程, policy gradient, actor-critic, SAC, PPO, 最大熵 RL, 探索/利用 (Haarnoja; OpenAI) |
ch12_行为克隆.md |
行为克隆 | BC, 复合误差/compounding error, 扩散策略 diffusion policy, 动作分块 action chunking (Chi et al.) |
ch13_视觉模仿学习.md |
视觉模仿 | UMI 手持夹爪, 跨具身 cross-embodiment, 相对轨迹表示, 鱼眼/隐性立体, 模仿学习数据 |
ch14_运动控制案例研究.md |
足式运动 | RMA 快速运动适应, 四足/双足行走, teacher-student 蒸馏, 域随机化行走, 盲走→视觉 (Kumar, Agarwal) |
ch15_导航案例研究.md |
导航 | GOAT, 语义地图, 可穿越性 traversability 估计, 自监督导航, 多目标导航 (Chang; Frey) |
ch16_灵巧操作案例研究.md |
灵巧操作 | Sim-to-Real 迁移, 域随机化, 阻抗/导纳控制, 触觉操作, 柔顺 manipulation, 域偏移分析 (Lin; Choi) |
ch17_长程规划与语言.md |
VLA / 长程规划 | VLA 视觉-语言-动作模型, π0.5, Molmo, 语言在规划中的作用, 长程任务, 通用具身智能 (Physical Intelligence; Kim) |
Cross-cutting topics (read multiple)
- Diffusion for policies: foundations in
ch04, applied to control inch12(diffusion policy) andch16(manipulation). - CPG / 运动基元: theory in
ch02, developmental view inch07. - Prediction & estimation:
ch09(Kalman/forward models) connects toch08(MPC) andch10(world models). - Sim-to-Real / domain randomization:
ch14(locomotion) andch16(manipulation). - Learning paradigms map:
ch11(RL) vsch12/ch13(imitation) — readch01for how the whole course fits together.
Answer style
- Match the user's language (the book is Chinese; answer in Chinese unless asked
otherwise), keeping key terms as
中文 (English)on first use, per the book. - Lead with the intuition, then the formal result, mirroring the book's "物理直觉 + 完整推导" structure.
- Cite the chapter and section so the answer is verifiable and the user can read more. When you pull an equation or algorithm, point to where it lives.