World Models & Spatial Intelligence
From Representation to Prediction, Planning and Physical Intelligence.
This is an open course on the intersection of world models and spatial intelligence: how agents learn a representation of the world from observations, how they predict the future inside that representation, how they plan and act on their predictions, and how all of this grounds out in 3D understanding, robotics and embodied intelligence.
Who this course is forโ
- Primary audience: graduate students and advanced undergraduates with a machine-learning background who want to enter world models / spatial intelligence / embodied AI research.
- Secondary audience: practitioners in LLMs, computer vision or robotics who want to systematically fill in the "world model" piece of the puzzle.
- Not for: absolute beginners (start with the external prerequisites listed in Foundations) or developers who only want to call a video-generation API.
How the course is organizedโ
The course is Y-shaped โ a shared foundation with two parallel exits:
| Track | Goal | Exit |
|---|---|---|
| ๐งฉ Foundations | Fill math/ML gaps, on demand | Jump back into either track at any point |
| ๐งโ๐ฌ World Model Scientist | Build world models: representation โ dynamics โ prediction โ planning โ evaluation | Reproduce Dreamer / TD-MPC-class systems |
| ๐ท Spatial & Embodied Engineer | Build systems: geometry โ 3D โ SLAM โ spatial memory โ navigation โ robots | Ship a perception โ state-estimation โ planning pipeline |
The two tracks are parallel, not sequential โ and they share the reliability layer (OOD, drift, evaluation) and the capstone.
Every module follows the same rhythm: why it matters โ the concepts โ curated references โ a hands-on lab on free Colab GPUs. Track your progress with the per-module checkboxes; the roadmap always shows you where you are.
Curated, not copiedโ
This course is a curator, not a mirror. Knowledge is carried by selected external links โ papers, official docs, and 11 surveyed university courses (see University Courses) โ while original effort goes into structure, connective text, roadmaps and notebooks. Third-party course materials are indexed by link only and are never redistributed.
Course content is licensed CC BY 4.0; lab code and notebooks are MIT.