What it is
Microduck Studio is a free, open source app for the Microduck, the small walking duck robot from Pollen Robotics. It runs on iPhone and iPad, and as a native Mac app from the same code. It is in an open TestFlight beta. It is not the official Microduck app and I am not affiliated with Pollen.
Nobody has the hardware yet, including me. Pollen’s first deliveries are expected around Christmas 2026. So the app is built for the robot people are waiting for, and everything it claims about that robot is measured in simulation first.
What you can do in it:
- Play and Learn. Drive the duck from a game pad, run its tricks, and follow a short learning path into the deeper tools.
- Policies. Load a trained network, see what is inside it, check whether it is one of Pollen’s, and blend or search new ones on a physics bench.
- Motions and plans. Write a motion keyframe by keyframe, or describe it and let a language model draft it, checked against the robot’s real joint limits. A plan editor chains them into a run.
- Duck Cards. A free deck of coding cards you print and cut out. A child lays Walk, Turn and Repeat cards in a row on a 30 cm grid and the duck acts the row out. Packs turn that into a sequence of challenges.
- Compare. Watch two runs side by side and pick the better one. Enough picks become a reward, and the reward trains a new policy on Hugging Face Jobs: reinforcement learning from human feedback (RLHF), steered from a phone.
- The Lab. Duck Soccer, a life size duck in AR, a live camera view from the robot, and evaluations that write a log somebody else can open and check.
How it works
The app is Swift, split into a UI free kit that computes and an app that only draws, so almost everything can be tested on a Raspberry Pi with no phone attached. The robot’s own truth, its joints and its policies, lives in a separate public package pinned by version.
The interesting part is Duck Cards, because it asks a question with a real answer. A timed card (“walk for two seconds”) never lands on a square: speed changes with the floor, the policy and the battery. A closed loop card walks until odometry says one square and turns until heading says a quarter turn. Is that good enough for a classroom?
I wrote down what “good enough” means before running anything, then ran thousands of card rows in Pollen’s own MuJoCo simulator and published every row at microduckstudio.com/science. The short answer so far is no, and the page says exactly where:
- Single cards mostly pass. Long rows of ten cards or more mostly did not.
- Four fixes were each pre-registered, run, and judged one at a time. Calibrating heading drift while standing was rejected. Using the floor mat as a compass lifted long rows a lot but was not adopted, because it added falls and its repeat missed its own validity gate by one row. Merging consecutive cards is kept, with a small gain. Nests and chapters, where a long challenge plays in chapters and the child re-seats the duck on a marked square between them, raised long rows from 42% to 63% and was kept. A fifth idea brought the mat back, laid under each nest, and let the mat measure the camera’s own tilt; it raised long rows from 63% to 97.5%, and every prediction written down before the run held. It is the new base in simulation, and not in the app yet: it needs the duck’s camera in the app and a printed mat.
- Running the same code twice gave seven different verdicts, because rows in one simulated session share a duck. So later fixes are judged by a permutation test over sessions with a Holm correction, not by counting rows.
Until the stressed condition passes, closed loop cards stay off on real ducks. In the shipped app Walk and Turn run as timed cards, and the paper board keeps the rule.
I built it pairing with Claude Code. The evaluations now run on Hugging Face Jobs rather than on my desk, so each one is a few thousand rows on a fresh machine with its logs attached.
Why it matters
A robot for kids should come with evidence, not a demo video. Most of what this app knows about the Microduck is a measured “not yet”, and I would rather ship that, with the rejected ideas next to the kept ones, than a feature that only works the one time it was filmed. When the ducks arrive, the same cards will be held to the same written bar.