Skip to main content

Hugging Face's $400 Microduck: A Tiny Robot You Can Train Yourself

Hugging Face's open-source robotics company, Pollen Robotics, has just dropped something that's equal parts adorable and impressive: a mini bipedal robot named Microduck. Standing just 25cm tall and weighing a mere 800g, this little machine duck is more than a desk toy—it's a fully open-source platform for diving into the world of reinforcement learning and embodied intelligence.

Image

Small Size, Big Moves

Don't let its size fool you. Microduck is packed with a Rockchip RK3566 processor and drives 15 servos at a 50Hz control frequency. It's already mastered a bunch of practical skills: it can walk on two legs using a game controller, stand up on its own after a tumble, pick up objects from the ground with its beak, and even perform tricks like sitting, standing, kicking a ball, and doing a forward roll. But here's the kicker—if you attach wheels to its feet, it switches to a different reinforcement learning strategy and can slide, turn, go downhill, and even stand up while roller-skating. The official team jokes, "No ducks were harmed in the making of this robot, although we consulted several ducks."

Almost Everything Is Learned, Not Programmed

The training project, microduck_rl, is open-sourced and follows a clear technical path. It uses MuJoCo / MuJoCo Warp for simulation, the PPO algorithm with domain randomization to bridge the sim-to-real gap, and then exports an ONNX model for direct deployment on the robot. The strategy runs at 50Hz, matching the real robot's control rhythm. To make the transition from simulation to reality smoother, the training environment even simulates motor friction, battery voltage changes, control delays, and servo gear backlash—so the learned behaviors are robust enough to work in the real world.

A Software Architecture That Means Business

The robot's software is written in Rust and split into several independent services, each with a specific job:

  • robotd: Handles real-time control, executes reinforcement learning policies, and ensures safety.
  • updaterd: Manages OTA updates, verifies signatures, and rolls back automatically if something goes wrong.
  • configd: Deals with Wi-Fi and device identity.
  • btd: Manages Bluetooth communication.
  • padd: Supports game controllers.
  • mediad: Handles camera, audio, and WebRTC.
  • tofd: Manages the ToF depth sensor.

These services talk to each other via Unix Socket and JSON-RPC 2.0. Only robotd has direct motor control permissions. Commands from your phone, a game controller, or custom code are just "action intentions"—they still have to pass robotd's safety checks before anything actually moves. That's a smart way to keep things reliable at the architectural level.

OTA Updates Without the Fear of Bricking

The update process is thoughtfully designed: download the release, verify the signature, switch versions, restart the service, run a health check, and if anything goes wrong, automatically roll back. So even if a new update causes problems, the robot won't end up as a paperweight.

More Than a Toy: A Miniature Embodied Intelligence Platform

Microduck is way more than a "walking machine duck." It crams the entire pipeline—reinforcement learning training, sim-to-real transfer, ONNX deployment, Rust real-time control, WebRTC audio-visual, and a complete OTA mechanism—into that tiny 25cm, 800g body. For developers and researchers interested in reinforcement learning, robot motion control, or embodied intelligence, this open-source project offers a complete reference from simulation training to real-world deployment. The entry barrier is low, but the potential is high. Both the runtime and the training code are open-sourced, so you can dive in and start experimenting today.

Key Points

  • Microduck is a 25cm, 800g open-source bipedal robot from Pollen Robotics (Hugging Face).
  • Costs under $400 and can be trained by users.
  • Uses reinforcement learning and sim-to-real transfer for skills like walking, self-righting, and roller-skating.
  • Software is Rust-based, modular, with safety-focused architecture.
  • Includes OTA updates with automatic rollback.
  • Open-sourced training code and runtime for developers and researchers.