Microduck Open-Sources Complete Sim-to-Real Reinforcement Learning Stack, Runs Policies via ONNX on a 25 cm Bipedal Robot

Pollen Robotics, a Hugging Face subsidiary, has introduced Microduck, a 25 cm-tall bipedal robot weighing about 800 grams. It features 15 motors, a camera, LiDAR, two IMUs, and Wi-Fi, Bluetooth, and NFC connectivity. The company has released more than just a high-level SDK: the on-device Rust software, MuJoCo models, PPO training environment, domain randomization, ONNX export pipeline, and factory-installed motion policies are all available on GitHub under the Apache 2.0 license.

The device is built around a Rockchip RK3566, with neural policies controlling the servos locally at 50 Hz. `robotd` handles the control loop and motor bus, while `mediad` streams camera video over WebRTC. Bluetooth, configuration, depth sensing, and updates are managed by separate daemons; all components communicate through JSON-RPC over Unix sockets. The update system includes signature verification, health checks, and rollback, making this a physical platform designed for sustained modification rather than merely a released demonstration notebook.

On the training side, MuJoCo Warp runs large numbers of environments in parallel. In the official example, training a usable gait across 4,096 environments takes roughly one to two hours, after which the policy is exported to ONNX for execution on the robot. The simulation randomizes battery voltage and voltage sag, command latency, friction, actuator back electromotive force, and gear backlash. The backlash-enabled model uses additional passive joints to simulate approximately one degree of play in either direction for each servo. Policies for different walking modes, fall recovery, sitting and standing, picking up objects, kicking a ball, and roller-skating share a 61-dimensional observation interface and can be hot-swapped at runtime.

Microduck’s value lies in offering a lower-cost, traceable research baseline spanning simulation through deployment on physical hardware—not merely in its cute appearance. However, “fully open source” does not mean the sim-to-real problem has been solved. The official demonstrations primarily cover predefined motor skills and have yet to establish stability on unfamiliar terrain, with aging hardware, after collisions, or during prolonged operation. The first units are still in the preorder stage. Researchers should next examine manufacturing variance across production units, whether policy interchange formats remain compatible, and whether the community can reproduce new skills without recalibrating the entire simulator.