fanhao375/microduck-replica

Microduck 复刻：飞特 HD-1910 整机已实机验证（站立、行走、前滚翻起身）。镜像、Studio 网页中控、CAD/BOM、装配教程、电路工程与训练平台 | Microduck robot replica with real hardware demos.

What is this project?

Microduck replica is a community‑driven effort to rebuild the Microduck robot – a 25 cm tall, 737 g bipedal “duck” robot that walks using reinforcement‑learning‑trained policies. The original robot’s software is open‑source, but many hardware files (CAD, PCB schematics) were not publicly released. This repository reverse‑engineers the official MJCF simulation model to produce:

  • Detailed assembly drawings and exploded views.
  • Editable CAD files (SolidWorks, STEP) for two actuator options – the original Dynamixel XL330 servos or the cheaper Feetech HD‑1910 servos.
  • A reconstructed PCB design for the previously closed‑source imu_to_dxl board.
  • Documentation, BOMs, and step‑by‑step video tutorials for building the robot from scratch.
  • Forks of the control software that add direct IMU‑to‑I²C support and a web‑based debugging UI.

In short, it turns the Microduck from a simulation‑only platform into a fully reproducible physical robot.

Who might use this?

  • Robotics hobbyists who want a compact biped platform for experimenting with locomotion, reinforcement learning, or sensor integration.
  • Researchers/educators needing a low‑cost, open‑hardware robot for labs or teaching labs.
  • Makers interested in 3‑D‑printing, PCB fabrication, and integrating custom sensors (IMU, depth camera) on a small robot.

Key features

Feature Details
Two actuator families • XL330 – the original Dynamixel servo (15 units, 1 kg·cm torque).
• HD‑1910 – Feetech’s cheaper, higher‑torque alternative (≈2.5× torque, half the price).
Hardware documentation • Fully open‑source RPI Robot HAT board (KiCad, Gerbers, BOM).
• Reconstructed imu_to_dxl board (firmware, register map, validation docs).
• Exploded assembly drawings, part mass/inertia data extracted from the MJCF model.
CAD files Editable SolidWorks and STEP assemblies for both servo versions, released via the companion microduck-replica-cad repo.
Software fork Based on the upstream Microduck‑kissqy controller (Apache‑2.0). Adds direct IMU‑I²C integration (BNO085/LSM6DSV16X) and a web‑based servo‑debug UI that visualises joint angles and robot pose in 3‑D.
Training pipeline Links to the original reinforcement‑learning environment (microduck_rl) and to the xgoduck_rl baseline. Provides scripts for data preparation, training, and ONNX model export (though full‑robot walking policies are still being validated).
Step‑by‑step guides Video series on B‑站 covering mechanical assembly, wiring, and initial power‑up; written docs for PCB assembly, firmware flashing, and IMU calibration.
Community support A WeChat group for builders, issue tracker for bug reports, and a list of related community projects (color studio, CNC‑reinforced parts, etc.).

How to get started

  1. Choose your servos – XL330 (original) or HD‑1910 (cheaper, higher torque). The README’s comparison table lists voltage, mechanical differences, and which CAD release to use.
  2. Gather parts – Follow the BOM.md for a complete parts list (servos, bearings, fasteners, NP‑F550 battery, Radxa Zero 3W board, IMU module, camera, etc.).
  3. Print or fabricate – Download the appropriate CAD release from microduck-replica-cad and print the 3‑D parts. Order the RPI Robot HAT PCB (open source) and, if needed, fabricate the imu_to_dxl board using the provided schematics.
  4. Assemble – Watch the B‑站 tutorial series (8 episodes) that walks through each mechanical sub‑assembly.
  5. Wire & flash – Connect the servos, IMU, camera, and HAT to the Radxa Zero 3W. Use the provided scripts to flash the imu_to_dxl firmware and the main robot controller.
  6. Run software – Clone the forked controller (Microduck‑kissqy), enable the IMU mode in robotd.toml, and launch the web‑debug UI to test joint movements and pose estimation.
  7. Train or load a policy – Use the training scripts in software/training/ to train a new walking policy, or load an existing ONNX model if your hardware matches the original configuration.

Why it matters

The original Microduck robot demonstrated that a small biped can learn to walk using reinforcement learning, but its hardware files were fragmented. By reconstructing the mechanical design, PCB schematics, and providing clear assembly and software instructions, this project lowers the barrier for anyone to build and experiment with a real‑world RL‑driven robot. It also showcases a clever hardware trick: the IMU pretends to be a Dynamixel servo on the same single‑wire TTL bus, eliminating the need for a separate I²C line and simplifying the controller architecture.

Quick links

  • CAD & STL repo – https://github.com/fanhao375/microduck-replica-cad
  • Software fork – https://github.com/fanhao375/Microduck-kissqy
  • Assembly videos – B‑站 playlist linked in the README
  • Documentation – docs/ folder contains hardware reverse‑engineering, IMU integration, and actuator selection guides.
  • Community – WeChat group QR (expires 2026‑10‑23) and issue tracker for support.

Related

  • Dispatch
    Pollen Robotics Microduck: An Open-Source Biped for RL Training

    Pollen Robotics has launched Microduck, a 25cm open-source biped robot designed for reinforcement learning (RL) training via a sim-to-real pipeline using MuJoCo.

  • Project
    AI-FanGe/Microduck-build-tutorial

    A compact biped robot project providing a full pipeline from MuJoCo-based RL training to real-world deployment on a Raspberry Pi Zero 2 W using ONNX policies.

  • Project
    pollen-robotics/microduck

    A control system for a tiny biped robot that uses reinforcement learning policies to walk, roll, and interact with its environment.

  • Project
    LuwuDynamics/xgoduck_hardware

    XGO-Duck is an open-source, 3D-printable bipedal robot platform designed to bridge the gap between physical hardware construction and reinforcement learning-based motion control.

  • Project
    menloresearch/asimov-1

    An open-source humanoid robot project providing the CAD, simulation models, and software needed to build and customize a 1.2m bipedal robot.