AI-FanGe/Microduck-build-tutorial

A practical hardware and software setup for a compact RL-powered biped robot.

What it solves

Microduck is a compact biped robot project that provides a complete hardware and software pipeline for implementing reinforcement learning (RL) walking policies on a physical robot. It bridges the gap between simulation training and real-world deployment by providing a pre-configured OS image, 3D-printable parts, and a deployment framework.

How it works

The project is split into two primary components:

  • Simulation & Training: Using mjlab_microduck (based on MuJoCo/MjLab), a walking policy is trained via reinforcement learning and exported as an ONNX model (walk.onnx).
  • Deployment: The microduck/ directory contains code for a Raspberry Pi Zero 2 W that reads data from an IMU (BNO08x) and user inputs (keyboard or Bluetooth gamepad), then executes the ONNX policy to drive 14 Dynamixel XL330 servos via an OpenRB-150 controller.

Who it’s for

  • Robotics enthusiasts and developers interested in bipedal locomotion.
  • Researchers or hobbyists wanting to experiment with RL-based walking policies on affordable hardware.
  • Users who want a "turn-key" experience via a pre-built OS image and 3D printing files.

Highlights

  • Complete Ecosystem: Includes everything from CAD files for 3D printing to the RL training environment and deployment code.
  • Pre-built Image: Provides a microduck.img.xz image with Raspberry Pi OS Lite, Python environments, and the walking model pre-installed.
  • Headless Operation: Features a dedicated systemd service allowing the robot to be started and controlled via a Bluetooth gamepad without needing a monitor or keyboard.
  • Hardware Integration: Built-in support for BNO08x IMUs and ROBOTIS OpenRB-150 servo controllers.

Related

  • Project
    iamaisim/ProjectAirSim

    An open-source, engine-independent simulation platform for autonomous systems that supports both lightweight headless execution and high-fidelity 3D environments via Unreal Engine 5.

  • Project
    ntnu-arl/gbplanner_ros

    A unified graph-based planning framework that enables aerial, ground, and underwater robots to autonomously explore unknown environments and inspect structures.

  • Project
    icey1287/SuperMew

    SuperMew is a self‑hosted, audit‑first RAG/agent platform. It persists threads, runs, events and document versions, offers hybrid Milvus+BM25 retrieval, supports HITL checkpoints, versioned skills/tools, a frozen model control plane, and a built‑in evaluation framework. The stack uses FastAPI, PostgreSQL, Redis, Milvus, and a Vue 3 frontend, and can be run locally with Docker‑Compose or deployed in production with systemd/K8s.

  • Project
    snape/RVO2-CS

    A C# implementation of the Optimal Reciprocal Collision Avoidance (ORCA) algorithm for enabling multiple independent agents to avoid collisions without communication.