Robot Learning: From Fundamentals to Foundation Models
- 提供者: ETH Zurich (Computer Vision and Geometry Group)
- 講師: Oier Mees
- 程度: advanced
- 形式: video-lectures
- 時長: 12 weeks, 5 ECTS (11 recorded lectures, Feb 16 - May 18 2026)
- 類別: robot-learning
- 語言: English
- 價格: 免費
- 標籤: robot-learning, imitation-learning, reinforcement-learning, vision-language-action, foundation-models, world-models, diffusion-policy, eth-zurich
ETH Zurich 2026 年春季的機器人學習研究所課程(263-5911-00L),由 Oier Mees 授課,完整內容公開在 Computer Vision and Geometry Group 的課程頁面上:十一份講義投影片並附 YouTube 錄影、十場客座演講錄影(講者包括 Cheng Chi、Ted Xiao、Quan Vuong、Aviral Kumar 與 Lucas Beyer)、每週精選的論文清單,以及放在公開 GitHub 儲存庫中的四份程式作業。課程內容從機器人控制與 MDP,經過模仿學習與強化學習,一路到生成模型、世界模型與通用的視覺-語言-動作策略。這些教材可免費閱讀與觀看,但儲存庫沒有授權檔案,課程頁面也未載明授權,因此應視為可供瀏覽,而非開放授權。
為何值得修讀
This is the most current graduate treatment of robot learning available for free, and currency is the point: a 2026 syllabus that spends whole weeks on world models, generalist policies and embodied reasoning, discussing papers published months before the lecture. Oier Mees works on exactly this material, and each week closes with a guest whose own system is on the slide - Cheng Chi on diffusion policy and UMI, Quan Vuong on Physical Intelligence's VLAs, Scott Reed on NVIDIA GEAR's world models. Where the older free courses here teach the mathematics and classical methods that stay true, this one maps where the field actually is.
你將學到
- Robot control and Markov decision processes as the formal frame for everything that follows
- Imitation learning: behavioural cloning, its failure modes, and visual representation learning for manipulation
- Reinforcement learning for robots across two weeks, from policy gradients to human-in-the-loop and offline RL
- Generative models as policies - diffusion policy, implicit behavioural cloning, and steering them with RL
- Sequence modelling and transformers for control: decision transformers, ALOHA-style bimanual policies, next-token locomotion
- World models and video-generation policies for learning behaviour without a physical robot
- Generalist Vision-Language-Action policies, from language-conditioned imitation through Gato to pi*0.6
- Embodied reasoning and test-time scaling, and the open problems the field has not solved
課程大綱
Twelve weekly sessions, February to May 2026. Weeks 1-11 each pair a lecture (slides plus YouTube recording) with a student-led discussion of three papers, and weeks 2-11 close with a guest spotlight recording. The arc: introduction, robot control and MDPs, imitation learning, reinforcement learning I and II, generative models, sequence modelling and transformers, world models, generalist robot policies, embodied reasoning and test-time scaling, frontiers and open problems. Week 12 is guest lectures only. Four coding assignments - PyTorch and NumPy, robot control and MDPs, imitation learning, reinforcement learning - live in the course GitHub repository, which holds the homework folders and nothing else. Enrolled students are graded on paper presentation (20%), homework (40%) and a group final project (40%); none of that is available to self-learners, who get the lectures, papers and assignment code.
核心主題
- Robot control and MDPs
- Imitation learning and behavioural cloning
- Reinforcement learning for manipulation and locomotion
- Diffusion and other generative policies
- Transformers and sequence modelling for control
- World models
- Generalist robot policies and VLAs
- Embodied reasoning and test-time scaling
先備知識
- Solid linear algebra, calculus and probability (the ETH catalogue's stated requirement)
- Python - all assignments are Python, and the first homework is a PyTorch and NumPy tutorial
- Background in machine learning and robotics is strongly encouraged rather than required
適用對象
- Graduate students and researchers who already know ML and want the robotics-specific state of the art
- Engineers building manipulation or locomotion policies who need the vocabulary behind current papers
- Anyone who has finished a classical robotics or RL course and wants the 2026 continuation
- Reading-group organisers - each week ships three papers with a recorded discussion