Modern Robotics and Physical AI
A course on how modern robots work: from classical control and behavior cloning to VLA models and humanoids. Every topic comes with labs in your browser, where small neural networks train in a few seconds.
How a lesson works
A lesson takes 30–45 minutes and covers one idea. It starts with a problem the previous methods failed to solve, then works through the idea and the key paper, and then lets you test it all hands-on.
Theory on two levels
The main text builds on intuition and diagrams. Formulas, dimensions and code live in collapsible blocks.
Hands-on labs
Experiments with small models right on the page. Before each run you make a prediction, afterwards you compare it with the result.
Working with papers
Each lesson builds on one or two key papers. A reading guide tells you what to read closely and what you can skip.
Checking your understanding
A quiz with explained answers and open questions with reference solutions.
Program
9 parts, 54 lessons. The parts follow a logical order: each one answers the limitations of the approaches in the previous one. 34 lessons form the core of the course; the other 20 are deep dives into specific topics.
Terms from the lessons with short definitions are collected in the course glossary. What has changed in the lessons is on the “What’s new” page.
The “completed” and “started” marks are stored only in this browser.
Who this course is for
The course is written for engineers with experience in one of two fields. You need Python and the basics of machine learning. Transformers, diffusion models and the basics of control theory are covered inside the course.
ML engineers moving into robotics
- kinematics, control and planning, as much as you need to work with learned models
- what makes robotics data special and how it is collected
- the problems where classical methods still beat neural networks
Roboticists picking up machine learning
- a step-by-step path from behavior cloning to VLAs and world models
- how π0, GR00T and Gemini Robotics are built
- ways to combine learned models with the classical stack
What you will be able to do
Break down the architecture, data and limitations of new robot learning papers.
Choose between learned, classical and hybrid solutions and justify the choice.
Estimate how much teleoperation, simulation and human video you need and how to set up a data flywheel.
Tell a demo from a finished product and understand the strategies of the companies.
Oleg Shipitko
I have worked in robotics for over ten years. This course is what I wish I could have read myself when neural networks came to robots.
Readers make the course better
Readers’ comments and suggestions make their way into the lessons. Thank you to everyone who writes: in the channel, in the feedback at the end of lessons, and on GitHub.