Hugging Face Microduck: $399 Open-Source Robot With Full RL Stack

Hugging Face opened preorders for the Microduck, a $399 open-source biped robot built by Pollen Robotics, which Hugging Face acquired in April 2025. The 25 cm, 800-gram robot ships with a complete reinforcement learning training pipeline on GitHub, including a sim-to-real stack calibrated to its Dynamixel XL330 servos, and preorders are scheduled to ship before Christmas 2026. The launch comes two days before Nvidia's $12.9B acquisition of Hugging Face became public, raising questions about future open-source commitments.

Hugging Face opened preorders this week for the Microduck — a $399 open-source biped robot built by Pollen Robotics, the French startup HF acquired in April 2025. At 25 cm tall and 800 grams, it has a beak that can pick up objects, rolls on skates, and falls down without breaking. It also ships with a complete reinforcement learning training pipeline on GitHub. The hardware is the packaging. The stack is the product. What You Are Actually Getting The robot itself is real hardware: 15 Dynamixel XL330 servos, a camera, LiDAR, dual IMUs, Wi-Fi, Bluetooth, and a Rockchip RK3566 running a 50 Hz control loop. Battery life is approximately one hour — the practical ceiling for experiments unless you grab the $39 charger pack. Four colorways Cream, Graphite, Lavender, Sky . Preorders ship before Christmas 2026, though first-production-run delivery targets on bipedal robots are targets, not commitments. The more interesting part is what Pollen Robotics published alongside the hardware: two public Apache-2.0 GitHub repositories. pollen-robotics/microduck is the runtime and SDK. pollen-robotics/microduck rl is the full reinforcement learning training stack, built on MuJoCo Warp mjlab with PPO. You train at 4,096 parallel environments on a CUDA GPU, and usable locomotion gaits come out in roughly one to two hours. The Sim-to-Real Pipeline Is the Real News Sim-to-real transfer is the hardest part of physical RL. Policies that work perfectly in simulation frequently collapse on real hardware because friction, motor delay, and sensor noise are never perfectly modeled. Pollen published the recipe that makes it work for Microduck — not as a paper, but as runnable code. The microduck rl repo includes a BAM actuator model for the XL330 servo: voltage control law, back-EMF, Coulomb and Stribeck friction terms. Domain randomization runs per environment across battery voltage, voltage sag, command delay, friction coefficients, and terrain. Backlash simulation is included. When training is done, policies export to ONNX via scripts/export.py , which bakes the observation normalizer directly into the graph. The runtime on the RK3566 loads that ONNX file and runs at 50 Hz. It works because the sim is calibrated to this specific hardware. That is not something most hardware startups ship on day one. The Price Context Pollen Robotics also makes Reachy 2, a full humanoid robot used by research labs at Cornell and Carnegie Mellon. Reachy 2 costs $70,000. The Microduck costs $399. That gap is not incremental progress — it changes who can do physical AI research. An ML engineer experimenting with locomotion policies no longer needs a university lab or a six-figure hardware budget. They need a CUDA GPU for training and a preorder slot. The open-source framing matters here too. Hardware schematics and software are both Apache-2.0. Community-contributed gait policies can live on the Hugging Face Hub alongside model weights and datasets already there. That flywheel is the reason HF acquired Pollen in the first place. The Timing Microduck launched two days before Nvidia’s $12.9B acquisition of Hugging Face became public. The acquisition has not closed. HF and Pollen are operating independently, and the open-source licenses on Microduck are already published. Whether Nvidia preserves that open-source commitment post-acquisition is an open question — but the code is out and the preorder is live. How to Get Started Preorders are open at pollen-robotics.com/microduck https://pollen-robotics.com/microduck/ . The $119 Dev Pack is worth considering if you plan to go deep on the RL workflow. The training repository is at github.com/pollen-robotics/microduck rl https://github.com/pollen-robotics/microduck rl — you can start training in simulation before the hardware arrives. There is also a browser-accessible simulator and a Discord community for questions. TechCrunch’s launch coverage https://techcrunch.com/2026/08/27/hugging-face-is-selling-a-cute-399-open-source-duck-robot-microduck/ has additional background on Pollen Robotics and the acquisition context. Physical AI research just got $399 cheaper. For a field that previously required a grant or a well-funded lab, that is a meaningful shift.