Hugging Face's open-source robotics company, Pollen Robotics, has officially open-sourced a mini bipedal robot called Microduck (Micro Duck). This 25cm-tall, 800g machine duck is not only cute in appearance but also integrates cutting-edge technologies such as reinforcement learning, Sim2Real, and real-time control. It has become a highly valuable miniature embodied intelligence development platform.

image.png

Mini Size, Big Skills

Microduck is equipped with a Rockchip RK3566 processor and drives 15 servos at a 50Hz control frequency. It has already mastered various practical skills: it can walk on two legs using a game controller; it can stand up on its own after falling; it can pick up objects from the ground with its beak; and it can perform actions like sitting down, standing up, kicking a ball, and doing a forward roll. More interestingly, by adding wheels to its feet, it can switch to another reinforcement learning strategy, enabling sliding, turning, going downhill, and even standing up while roller-skating.

The official humorous message says: "No ducks were harmed in the making of this robot, although we consulted several ducks."

Almost All Actions Are 'Trained' Using Reinforcement Learning

The training project for Microduck, microduck_rl, has been open-sourced. The overall technical approach is clear: based on the MuJoCo / MuJoCo Warp simulation environment, the PPO algorithm combined with domain randomization is used to complete the Sim2Real transfer, then export an ONNX model, and deploy it directly on the real robot runtime. The strategy is trained at a 50Hz frequency, maximizing the match with the real robot's control rhythm.

To reduce the gap between simulation and reality, the training environment simulates motor friction, battery voltage changes, control delays, and even the backlash of servo gears, significantly improving the stability of the strategy's deployment.

Hardcore Software Architecture, Service-Oriented Splitting

The robot's software is mainly written in Rust, without relying on large frameworks, but instead split into multiple independent services:

  • robotd: Real-time control, execution of reinforcement learning strategies, and safety control
  • updaterd: OTA updates, signature verification, and automatic rollback in case of failure
  • configd: Wi-Fi and device identity management
  • btd: Bluetooth communication
  • padd: Game controller support
  • mediad: Camera, audio, and WebRTC
  • tofd: ToF depth sensor

Services communicate through Unix Socket + JSON-RPC 2.0. Only robotd has actual motor control permissions. Instructions from mobile phones, game controllers, or custom programs are just "action intentions," and final execution must pass robotd's safety verification, ensuring reliability from the architecture level.

Complete OTA Updates, No Fear of Bricking

The software update process is well-designed: download Release → verify signature → switch version → restart service → health check → automatic rollback in case of error. Even if the new version causes problems leading to the robot malfunctioning, the system can automatically revert to the previous version, effectively preventing the device from being "bricked."

More Than a Toy, a Miniature Embodied Intelligence Platform

Overall, Microduck is far more than a "walking machine duck." It packs all the key capabilities—reinforcement learning training, Sim2Real transfer, ONNX deployment, Rust real-time control, WebRTC audio-visual, and complete OTA mechanisms—into this small body that is only 25cm tall and weighs 800g.

For developers and researchers interested in studying reinforcement learning, robot motion control, and embodied intelligence, this open-source project provides a complete reference pipeline from simulation training to real robot deployment, combining low entry barriers with practicality. The related Runtime and reinforcement learning training code have both been open-sourced, and interested developers can explore further at the official repository.