pollen-robotics/microduck

The problem

Hobby and research robots that walk usually ship either as closed black boxes or as bare kits that leave you to write the whole motion stack yourself. Getting a cheap legged platform to actually stay upright, recover from a shove, and accept field updates without bricking is where most of the effort goes. This repo is the on-robot software for a ~25 cm, 800 g duck: the 50 Hz control loop that drives fifteen servos from neural policies, plus radios, camera, depth sensor and an update system. The policies themselves are trained in a companion repository, microduck_rl, and exported to ONNX that this code loads.

What's different

The common path is either a commercial quadruped with a vendor SDK, or a hand-rolled stack on a hobby servo controller with hand-tuned gaits. Here the behavior comes from reinforcement-learning policies trained in MuJoCo with PPO and a sim2real recipe, then loaded as ONNX on a Rockchip RK3566. The second distinguishing choice is the software architecture: several small Rust daemons over one JSON-RPC contract on Unix sockets, so the gamepad, the phone app, the console and your own script all speak the same calls. Third, updates are treated as a first-class problem — signed, health-gated and reversible. The README is unusually explicit about design rationale and open problems, which is rare for a hardware-adjacent repo.

Who it's for

People who already have or plan to buy the duck and want to modify behavior, drive it from a laptop, or build on its daemons. Robotics students and researchers who want a small, affordable legged platform with real sim2real tooling and a readable architecture. It is not for anyone wanting an out-of-the-box product to unbox and ignore, nor for teams that need production-grade SDK stability and long-term API guarantees — this is a young project tied to one specific piece of hardware. If you just want gait-planning theory, the RL training code lives elsewhere.

Cost of adoption

You need the physical robot; the README offers a simulator so you can run the real daemons against a MuJoCo body without one, but that is a development aid, not the point. The stack targets a specific ARM board, so expect embedded cross-compilation and ssh-based deploys. Rust throughout, Apache-2.0, no exotic license constraints. The README does not spell out install commands or host toolchain requirements in detail — check CONTRIBUTING.md and the dev-board setup page before committing. Maturity is the real risk: the update machinery, JSON-RPC contract and docs are all in motion, and hardware-coupled projects age quickly when the board revision changes.

Verdict

The repo is 87 days old and pulled in 10,859 stars with a single day on the trending board; that is a spike, not a track record. The +936 weekly Rust figure and a Hacker News thread with three comments suggest curiosity far outstripping usage. It is a well-organized, unusually candid codebase for a robot this cheap, and the design docs and simulator lower the barrier to experimenting. But its value is inseparable from owning the hardware, and the project has had no time to prove that the update path, the daemons and the docs survive a year of real robots. Worth a look if you have the duck or want to learn how a small legged robot's software is actually structured; wait if you need something to depend on.