XGO-Duck is an open-source, duck-shaped biped robot that combines printable mechanisms, joint control and audio/video interaction. Arduino UNO Q and 15 serial bus servos control the legs, neck, head and mouth. Walking and skating modes support robot-building projects, Python application development and reinforcement-learning experiments.
This manual covers the complete audio/video kit, including the camera, microphone, two speakers, fill light and ToF ranging module. Print the shell and structural parts and assemble the robot yourself. Once assembled, use a phone or computer browser to operate Duck; it has no onboard screen.
System architecture
Arduino UNO Q combines two computing environments on one board. Linux runs the web interface, motion-policy inference and audio/video and AI voice services. The STM32U585 microcontroller communicates with the servos and reads sensors. Arduino Bridge carries commands and feedback between them.
The diagram shows functional relationships. The expansion board distributes power, connects the servos and carries the IMU. ToF connects to the UNO Q Qwiic connector on a separate I²C bus. For physical connectors and cable orientation, use the numbered wiring photograph.
Hardware and joint layout
| Part | Configuration | Purpose |
|---|---|---|
| Computing and control | Arduino UNO Q, 2 GB RAM; Linux + STM32U585 | Python applications, policy inference, servo communication and sensor acquisition. |
| Legs, head and mouth | 15 HD1910M servos; 12 kg·cm; magnetic encoder feedback | Five joints per leg, four neck/head joints and one mouth joint. Policies output 14 joint targets; the mouth is controlled separately. |
| Camera and sound | 2 MP 1080P camera, microphone, two speakers and fill light | Browser video, photographs, voice input and audio playback. Runtime settings determine the actual video stream. |
| Attitude and ranging | QMI8658A six-axis IMU; VL53L5CX 8 × 8 ToF | Acceleration and angular velocity for state estimation; forward distance measurements. |
| Mechanisms and motion | V5 printed parts, soft foot pads and passive skating wheels | Walking / skating; skating uses joint motion with unpowered wheels. No powered wheel drive or robotic arm. |
| Power | 8.4 V battery pack and V1.1 expansion board | Power distribution and connectors for the controller, actuators and peripherals. |
See specifications for full component parameters and the kit and printing chapter for quantities and preparation.
What Duck can do
| Capability | How you use it |
|---|---|
| Biped motion | Use a browser or gamepad to walk, turn and run actions. Fit the matching passive wheels before using skating mode. |
| Audio/video interaction | Watch the camera, take photos, control the fill light, test the speakers and listen to the microphone. |
| AI voice | Connect to the internet, bind an account and select an agent; wake Duck to talk or request actions subject to its current control state. |
| Ranging | Read the 8 × 8 depth grid and use valid measurements to limit forward speed. This is not path planning or cliff detection. |
| Open-source development | Modify Python applications, study the Arduino/STM32 controller or train policies on a CUDA-capable computer. |
How movement and voice work
From a forward command to joint movement
A browser or gamepad sends a travel or turning command. Linux handles it according to the active mode and sensor state.
The onboard ONNX policy runs at 50 Hz, combining the command with IMU and joint feedback to produce 14 joint targets. The mouth target is handled separately.
Bridge sends targets to STM32. Its 100 Hz servo bus loop communicates with all 15 servos.
Servo and IMU feedback return to Linux for the next cycle. Valid ToF data can limit forward speed; it does not plan a route.
From wake word to spoken response
The microphone supplies audio to a separate Linux voice process. After wake-up it connects to the voice service selected by the bound account, and the speakers play the response. Action requests pass through the same motion-control checks, including mode, calibration and feedback. The media service supplies camera images to the cockpit.
Local browser control and motion inference run on Duck; access them over your local network. AI voice conversation needs internet access, account binding and an agent. Train new policies on a development computer, then deploy the exported ONNX model to Duck.
Continue with system architecture and Python development or policy training and export.
First build: follow these seven steps
Install the software and assign servo IDs before assembling the body. Each chapter includes instructions and completion checks; skip steps you have already completed.
01 · Check the kit and print the parts
Check the packing list, tools and complete-kit V5 parts.02 · Connect the board and find its IP
Complete first-time setup and record the Linux login.03 · Install and start XGO Duck AI
Open the cockpit and motion control console.04 · Set up and number the servos
Configure and label all 15 servos, one at a time.05 · Assemble the robot and wire audio/video
Build the body, fit the head modules and check every connection.06 · Calibrate and take the first steps
Calibrate the IMU and joints, then check the standing pose.07 · Bind, talk and verify the complete robot
Bind an agent and check sound, images, ranging and motion.
Already assembled? Find the task you need
Software packages, MakerWorld model pages and source projects are collected under downloads and resources.
Wiki