Microduck vs Open Duck Mini: Buy or Build?

The Short Answer

Buy Microduck if you want to start training policies this month, and build Open Duck Mini if you want a robot you can repair and redesign. Every buy-or-build comparison we publish here starts with the parts list, and on these two the money comes out close to a wash.

Microduck is $399 assembled, from Hugging Face and Pollen Robotics. Open Duck Mini v2 is a fully open project by Antoine Pirrone targeting a bill of materials under $400, published on GitHub with print and assembly guides.

Both are trained the same way. Both use MuJoCo for simulation and reinforcement learning for locomotion, and both put serious work into sim-to-real transfer, so the skill you build on one carries to the other.

The Parts Cost Is About the Same

A $399 assembled robot against a sub-$400 parts target means self-building saves you almost nothing in cash. That is the single most useful fact in this comparison, and it reverses the usual DIY assumption.

The reason is that both machines need the same expensive things: a dozen or more servos, a compute board, an inertial measurement unit (IMU), a battery, and a camera. Buying those individually at retail costs roughly what a manufacturer pays at volume plus their margin, once you account for shipping several orders.

DIY buys you control. It does not buy you savings. You choose the servos, you can upgrade one subsystem without replacing the robot, and nothing is sealed against you.

There is also a cost that never appears in either figure. A 3D printer, or a print service, is assumed by Open Duck Mini, and a CUDA graphics card is assumed by both training stacks. Neither project hides this. It just does not appear in a headline price.

What You Get for the Build

Open Duck Mini gives you every design file, and Microduck gives you none. That is the clearest difference between the two and the one that will still matter in two years.

The Open Duck Mini mechanical design is authored in Onshape, so you can open the model, change a bracket, and print the new part the same evening. Locomotion is trained end-to-end with reinforcement learning, and the project documents itself as it evolves, with a bill-of-materials spreadsheet, print guides, and assembly guides.

Microduck publishes its SDK, its simulator, and its full reinforcement-learning training stack under Apache-2.0, and stops there. The mechanical and electronic designs are not released, so a snapped leg has no published fix beyond contacting Pollen Robotics.

The two projects are not strangers. Open Duck Mini lists Hugging Face and Pollen Robotics among its sponsors, and both take the same MuJoCo-plus-reinforcement-learning route to a walking gait. Neither company has described Microduck as derived from it, so treat them as parallel projects that share an approach.

Where First Builds Fail

Most first attempts stop during the build. A bipedal robot has failure modes a wheeled one does not, and each of them costs a weekend to diagnose.

Servo horn alignment is the classic one. Assemble a leg a few degrees off its zero position and the trained policy inherits a bias it never saw in simulation. The robot then walks in a curve, or falls consistently to one side. Nothing in the code is wrong, so the debugging goes in the wrong direction for hours.

The second is the sim-to-real gap on friction and mass. A policy trained against slightly wrong inertia or a slippery foot works in MuJoCo and shuffles on your kitchen floor. That is why the Open Duck Mini project treats sim-to-real as its own workstream rather than a final step.

Microduck removes both problems. It ships calibrated, with policies already transferred to that exact hardware. That transfer work is the real product, and it is why $399 is not a markup on a parts list.

Spares are where the two diverge again over time. Pollen Robotics sells three spare Microduck motors in a $119 Dev pack, which covers the commonest failure on a servo-driven biped. A printed Open Duck Mini leg costs filament and an hour of printer time, and a replacement servo costs whatever the open bill of materials specifies.

What the Licences Let You Do

Only Open Duck Mini lets you fabricate and sell a derivative of the physical robot. It is a narrow point. It also decides the choice for anyone building a product, a course kit, or a lab fleet.

Open Duck Mini publishes its design files openly. A university can print thirty of them for a robotics module without asking anyone. The bill of materials is public too, so a department can budget the exercise before committing.

Microduck's Apache-2.0 licence covers the SDK, the simulator, and the training stack. For software it is genuinely permissive. Fork the training code, ship a product built on it, keep your changes private. It grants nothing over the hardware, because the mechanical and electronic designs are not published.

The practical test is simple. Ask whether your project needs to change the software or the machine. Software-only projects lose nothing by buying Microduck. Anything that touches the physical robot needs the open files.

A Ten-Minute Way to Decide

Five questions settle this faster than a spec table. Answer them about the next three months rather than about the person you intend to become.

• Do you own a 3D printer or have easy access to a print service? No means Microduck.
• Have you assembled a multi-servo robot before? No means the Open Duck Mini build will take longer than you expect.
• Is the interesting part the gait policy or the machine itself? Policy means Microduck.
• Do you need more than one unit, for a class or a lab? Open files and a public bill of materials scale better than a single retail product.
• Would a broken leg end the project? If yes, you need the printable design files.

What Each Community Gives You

Open Duck Mini's support is a Discord server and a documented repository, and Microduck's is Pollen Robotics. Those fail in different ways.

The open project's documentation evolves alongside the hardware, which means a guide can be a version behind the design and you resolve the mismatch by reading the commit history. That is normal for a live open project and frustrating on a deadline.

Pollen Robotics answers the question or does not, and at three days old Microduck has no accumulated public record of anyone else hitting your problem. Its sibling Reachy Mini took months to build that up.

Neither situation is a reason to avoid either robot. It is a reason to buy the one whose failure mode you can tolerate on a Sunday afternoon.

One practical hedge covers both. Run the simulator before you spend anything. MuJoCo is free and both projects train in it. An evening will tell you whether reward tuning is work you enjoy or work you resent.

Who Should Choose Which

Choose Open Duck Mini if you own a 3D printer, enjoy mechanical work, and want a platform nobody can close. It is also the better classroom project, because a broken part is a reprint rather than a dead robot, and thirty students breaking thirty parts stays a filament cost.

Choose Microduck if your interest is the policy and not the plastic. Researchers, students on a deadline, and developers who want to spend their weekend on reward functions all get more from a machine that already walks.

Skip both if you want a robot that does household work or entertains a child. Neither is a consumer product, and our robots for kids guide and home robot assistant guide cover those jobs properly.

What would change our answer is Microduck publishing its mechanical files or a price for structural spares. Either one would remove the repairability argument that currently carries the DIY route, and at that point the assembled robot wins outright for most buyers.

For the other direction, our Microduck versus Reachy Mini comparison covers the choice between the two Hugging Face products, and our open source robots guide covers the other buildable machines.