Microduck: AI-trained robot duck twins debut in flight simulator

A 25 cm open-source robot duck walks in a way you train yourself? That’s Microduck, and yes, it’s exactly as weird as it sounds.

What makes this a little more than just another cute robot isn’t the fact that it waddles around after you’ve pounded some reinforcement learning into it—it’s that you can upload its gait to a community of people who might want to make it do something even sillier. The launch simulator is already live, the policy-sharing button is there, and if you’re feeling generous (or just have an afternoon to waste), you can pre-order one while you’re at it.

The real question isn’t whether it’s impressive. It’s whether you can stand to look at a duck robot in your home for more than five minutes.

Technical Overview

The robotics stack in this project isn’t trying to reinvent the wheel, but it’s doing something most consumer robotics stacks don’t: shipping a sim-to-real pipeline that’s actually usable without a PhD. The key difference isn’t in the simulation software itself, but in how it integrates with the hardware controls and calibration tools.

The pipeline starts with a controller that can run in three modes: simulation, hardware-in-the-loop, and direct robot control. Switching between them is just a CLI command away:

$ robotctl monitor  # streams telemetry in sim or on real hardware
$ robotctl configure --mode=sim  # sets up the simulation environment
$ robotctl update --calibrate  # pushes updated PID gains to the robot

The sim-to-real transfer works because the controller abstraction is identical whether the robot is a URDF in PyBullet or a physical unit on your desk. The only thing that changes is the underlying driver — and even then, the interface stays consistent. That’s not magic; it’s just careful layering. Most robotics stacks either lock you into simulation-only or real-hardware-only workflows. This one doesn’t.

The real trick is the calibration layer. The robot’s joint offsets, gear ratios, and current limits are stored as structured config files. When you run robotctl update, it compares the simulated joint trajectories against the real ones and adjusts the hardware parameters automatically. It’s not perfect — backlash and stiction introduce error that PID tuning alone can’t fix — but it’s close enough for many applications. Close enough to make simulation useful, which is more than I can say for most open-source stacks.

The hardware side is where things get interesting. The robot uses a custom motor controller with deterministic real-time firmware, but the main CPU runs Linux. That split matters because the firmware handles the low-level current control while the Linux side manages path planning and higher-level behaviors. The latency between them is fixed at 2ms, which is tight enough for most hobbyist applications but not industrial-grade. If you’re trying to build something that needs sub-millisecond control, you’ll need to recompile the firmware — and that’s where the stack starts to show its limits.

Industry Impact

I don’t know whether to take the duck simulations seriously or not. A robot that behaves like a duck but isn’t a duck sounds like a toy, and toys aren’t the part of robotics where the industry stakes its reputation. Yet when you watch the clips, the motion is uncanny—the waddle, the head turns, the way it reacts to disturbances in the water—more like a biological system than a programmed sequence. If this is the same team that built the dog simulator, they’ve quietly shifted from quadrupedal physics to fluid dynamics, which implies a broader ambition than pet substitutes.

The battery-life complaints aren’t just griping; they’re a structural ceiling for anything that lives on your floor or pond. The community meme about the Aibo regret is shorthand for “we’ve already spent this money before,” which means adoption won’t hinge on cuteness alone. What changes the calculus is the simulator launch—if it lets researchers stress-test behaviors without buying hardware, it lowers the barrier to iteration. But I’m not convinced the industry will reward elegance in simulation with real-world investment, not when the next lever is cheaper motors or better battery chemistry.

I expect the first wave of adopters to be educators and researchers who need a controlled environment, not pet owners. For everyone else, the question isn’t whether the duck moves well—it’s whether the underlying stack is reusable for less whimsical robots. If the simulator becomes a Trojan horse for quadruped or drone control pipelines, then the duck is a footnote. If not, it’s another niche demo. The real tell will be the next project that ships with a “trained on our duck simulator” line.

Conclusion

Microduck isn’t just another toy with a jump-cut demo reel—it’s a real robot you can train yourself, not one more cloud API in drag. The fact that you can download the behavior policies and run them on actual hardware is still rare enough to matter, even if the duck gait is more "determined toddler" than "flight sim ace." At 25 cm and open-source, it’s small enough to break, big enough to matter.

I still don’t know what to make of the community policies. If people start uploading behaviors that turn the duck into a surveillance bot or a tiny drummer, Microduck’s creators will either look prescient or reckless. Either way, the real test starts when the first weird policy surprises them too.