Microduck Studio books
Four books, one little robot
A trained robot is easy to watch and hard to explain. These four books explain how a Microduck feels the floor, what its brain does with what it feels, how thousands of pretend ducks teach it to walk, and why a trained duck walks the way it does. Each book is ten lessons, and every lesson is a spread, with words on the left and a pop-up on the right that you play one step at a time.
Meet the Microduck
Senses, a brain and a loop; Duck Cards that make programs; why walking is hard; learning by trying; practice worlds; kind company in Kith; inventing a move.
For readers 7 and up, with a grown-up. No code.
How a Microduck Learns
The numbers it senses, the motors it drives, its small network, rewards and their loopholes, training at scale, PPO, domain randomization, and the road to a robot.
For teens and adults. Never writes code; shows the real numbers.
Training Microducks
Returns and discounting, advantages, the clipped objective, the observation contract, reward engineering, the motor model, measuring honestly, data from people, and shipping a policy.
For people who train robots, or want to, and can read a little algebra.
How a Policy Became
Real training runs, traced from their notes and logs: a walk that learned to stand, a kick taught to go straighter, and what a sensor on the duck's spare Qwiic port, or another simulator, would take.
For readers of Book 3, and anyone who wants to know why a trained policy does what it does.
Predict, play, check.
Every lesson asks for your prediction before it shows you anything. Then you press the steps on the right-hand page one at a time, with only the next one ready, and when the last step lands the book tells you whether you were right, and why.
Every number has a source
The books' numbers come from Pollen Robotics' own training and robot code, pinned to the version the books were written from, from this site's own pages and Kith, and in Book 4 from the notes and records of real training runs. Each lesson's notes say where, and before every deploy a check rereads each number's quoted source to make sure it still says so.
Toys are labelled
Some pop-ups are toy models: a tiny maze, a hill of scores, a curve drawn for the lesson. Each of those says so on the page, and the ones built from real settings say that instead.
Works without the pop-ups
Each book is also an ordinary page that you can read, search and print. A narrow screen gets one lesson at a time, with its pop-up in a card, and on a keyboard the arrow keys turn the pages and the space bar plays the next step.
With Duck Cards, Studio and Kith.
The books are where the ideas are explained, and the rest is where you try them: Duck Cards on a table, Microduck Studio on an iPhone, iPad or Mac, Kith, a web game by this site's author, and Pollen Robotics' microduck_rl on a computer with a graphics card. Every lesson ends with something to try in one of them. To train a duck of your own, Train your first duck takes you there step by step from a Windows or Linux PC, a Mac, or an iPhone or iPad, and Race in Pollen's Arena takes what you train to Pollen's own races on Hugging Face.
The duck in the pictures is our own drawing. It is a paper toy built with the robot's joints in the robot's order, not a model of Pollen Robotics' design.