Open Source Robotics

精選高品質的開源機器人專案、工具與軟體,推動機器人社群向前發展。

Principles of Robot Autonomy

  • 提供者: Stanford University (Department of Aeronautics and Astronautics)
  • 講師: Marco Pavone, with Daniele Gammelli, Joseph Lorenzetti, Katie Luo and Gioele Zardini
  • 程度: intermediate
  • 形式: Textbook, video lectures and notebooks
  • 時長: Self-paced: 19 chapters, 40 notebooks and 33 recorded lectures (14 in AA274A, 19 in AA203)
  • 類別: Robotics
  • 語言: English
  • 價格: 免費
  • 標籤: robot autonomy, motion planning, control, trajectory optimization, state estimation, SLAM, computer vision, reinforcement learning, imitation learning, model predictive control, python, stanford

Stanford 的機器人自主課程系列,以免費線上教科書的形式公開,並附上配套的講座、投影片與程式碼。這本書是 Daniele Gammelli、Joseph Lorenzetti、Katie Luo、Gioele Zardini 與 Marco Pavone 所著的 Principles of Robot Autonomy,正由 Cambridge University Press 出版,可在 porabook.com 免費閱讀全文,或在 arXiv 上取得 531 頁的 PDF。全書 19 章分為四個部分,涵蓋完整的自主堆疊:運動規劃與控制、感知、定位與建圖,以及決策,從動力學與軌跡最佳化,經卡爾曼濾波器與 SLAM,到強化學習與模仿學習。此外還有 40 個可在 Google Colab 中開啟的 Python 筆記本、Stanford Online 上兩門 Stanford 課程的錄影(AA274A Principles of Robot Autonomy I,錄製於 2019 年;AA203 Optimal and Learning-Based Control,錄製於 2026 年春季),以及兩門課程的 PDF 與 PowerPoint 講義投影片。

為何值得修讀

Most robotics teaching splits the field in two. Classical textbooks cover dynamics, control, planning and state estimation but were written before deep learning became part of perception and decision making, while machine-learning courses teach learning methods with little of the control and estimation that decide whether a learned policy works on hardware. This course puts both into one book with one notation, built around how the pieces fit together in a working autonomy stack. It is also unusually complete for something free: a full textbook, not lecture notes; Python notebooks for most chapters that run in Colab with nothing installed; real Stanford lectures rather than a recorded summary; and the slides an instructor needs to teach it. Every chapter also has Ask the Book, an optional assistant that answers questions from the chapter you are reading, using your own model account or a model run locally in the browser.

你將學到

  • Model robot dynamics, including nonholonomic wheeled robots, and simulate them numerically
  • Plan and optimize trajectories, then track them with feedback control: PID, LQR, iterative LQR and model predictive control
  • Find paths with graph search (A) and sampling-based planners (RRT, PRM)
  • Model and calibrate cameras, match features, and recover structure from stereo and motion
  • Detect and recognize objects with classical methods and deep networks
  • Estimate state with Bayes, Kalman, extended and unscented Kalman and particle filters, and localize a robot with them
  • Build maps while localizing, with EKF SLAM and pose-graph SLAM, and fuse several sensors
  • Make decisions with finite state machines, dynamic programming, reinforcement learning and imitation learning

課程大綱

The book runs in four parts. Part I, motion planning and control, covers modelling robot dynamics, open-loop control and trajectory optimization, closed-loop control and tracking, and motion planning. Part II, perception, covers sensors, camera models and calibration, stereo vision and structure from motion, classical perception, deep learning architectures and object detection. Part III, localization and mapping, covers filtering, approximate filters, robot localization, SLAM, and sensor fusion with object tracking. Part IV, decision making, covers finite state machines, dynamic programming, reinforcement learning and imitation learning. The 14 AA274A lectures follow Parts I to III. The 19 AA203 lectures go deeper into optimal control and learning: optimization theory, indirect and direct methods, dynamic programming and LQR, reachability, MPC, and reinforcement learning. The AA274A recordings are from autumn 2019 and are still being released. Homework, exams and solutions are not public; instructors can ask the authors for them.

核心主題

  • Robot motion planning and control: dynamics, trajectory optimization, closed-loop control, motion planning
  • Robot perception: sensors, camera models and calibration, stereo vision and structure from motion, deep learning for perception, object detection
  • Localization and mapping: Bayesian filtering, approximate filters, robot localization, SLAM, sensor fusion and tracking
  • Decision making: finite state machines, dynamic programming, reinforcement learning, imitation learning
  • Optimal and learning-based control (AA203): calculus of variations, Pontryagin's principle, direct methods, reachability, MPC, model-free and model-based RL

先備知識

  • Linear algebra, multivariable calculus and basic probability
  • Programming in Python; the notebooks use NumPy and, in places, JAX
  • Some exposure to differential equations helps for the dynamics and control chapters
  • No robot hardware: every exercise runs in a notebook

適用對象

  • Undergraduate and graduate students who want one text covering the whole autonomy stack
  • Controls engineers moving into learning-based methods, and machine-learning engineers who need the control and estimation side
  • Instructors building a robotics course, who can reuse the chapters, slides and notebooks
  • Self-taught engineers who want university material with working code alongside it

連結

  • 網站: https://porabook.com/
  • GitHub: https://github.com/StanfordASL/pora-exercises

本網站需要 JavaScript 才能正常運作。請啟用 JavaScript 以瀏覽我們的開源機器人專案。