What Open Duck Mini V2 Publishes About a Low-Cost Walking Policy

A small biped with the files attached

Open Duck Mini v2 is an open project led by Antoine Pirrone, aiming at a miniature of Disney’s BDX droid that a reader can actually build. The v2 README says the robot is about 42 cm tall with its legs extended, and that the full bill of materials should be under $400. Those two sentences are the hardware specification the project page itself is willing to print. [1] [2]

The same README calls the repository a working hub, with undocumented scripts, and says documentation is being written in pieces. The mechanical design is described as mostly settled, while expression parts (LEDs, a camera, a speaker and a microphone) can be added later. A walking duck in this repository is not yet the talking, looking character from the films. [1]

Training and the onboard computer are separate repositories

Walking policies are trained in simulation. The README says the project is moving that work to MuJoCo Playground and points to Open Duck Playground. Policies then run on a Raspberry Pi Zero 2W through Open Duck Mini Runtime, a third repository. CAD lives in an Onshape document, and the parts list is a spreadsheet linked from the README, with an English copy and a Chinese copy. [1] [3] [4]

That split matters for anyone reproducing the walk. The training code is not the binary on the Pi, and the Onshape model is not a calibrated MuJoCo file. The README warns that its guide from Onshape to MuJoCo has not been kept current. Following it without checking the Playground repository is a good way to simulate a robot the hardware no longer matches. [1]

What this page will not invent

Secondary write-ups list servo models, a joint count and a mass. The v2 README does not. Repeating those numbers here would pretend the project page had published them. Until they appear in the repository or in a paper with a method, they stay off this article. The under-$400 line is likewise a target in the README, not an invoice checked part by part on 13 October 2026. [1]

Walking clips are not a get-up result

The README’s sim-to-real section points at videos of the physical duck, including a push, and at policies a reader can try. It does not report a success rate, a speed, a surface, or a trial count. It also does not describe a policy that stands the robot back up after a fall. A viewer comment that they had not seen a get-up is consistent with that silence. It is not a measurement either. [1] [2]

The Pi Zero 2W is a small computer for a neural policy. The project chose it so the controller travels inside the robot. That choice leaves little spare compute for a second model, a camera stream or a high-rate estimator. The README does not publish the policy’s inference time, so the margin is unknown. What is known is the division of labour: MuJoCo for training, the runtime repository for the robot. [4] [1]

How a simulated walk is supposed to be checked on hardware is covered here. Sim-to-real for humanoid robots