AI in Practice

You can buy everything from online stores such as AliExpress and Amazon. Prices are rough ranges as of October 2026 and change with the seller, shipping and exchange rates. The total stays under $100, so spend what is left on spare wires and parts.

Connect everything with the power off. The microphone and amplifier pins match the ESP32-S3-CAM breadboard setup in the xiaozhi firmware, so you can use them as they are. The camera just plugs into the board’s flat connector. Tie all the GNDs together.

• Microphone INMP441 I2S

  WS

  GPIO 1

  SCK

  GPIO 2

  SD

  GPIO 42

  VDD

  3V3

  L/R · GND

  GND

• Amplifier MAX98357A I2S

  DIN

  GPIO 39

  BCLK

  GPIO 40

  LRC

  GPIO 41

  VIN · SD

  5V

  GND

  GND

• Motor driver DRV8833 / TB6612FNG

  AIN1

  GPIO 21

  AIN2

  GPIO 47

  BIN1

  GPIO 14

  BIN2

  GPIO 38

  VM

  7.4V (18650 ×2)

  GND

  GND

• Voltage converter 5V 3A buck converter

  IN

  7.4V (18650 ×2)

  OUT 5V

  ESP32-S3 5V · MAX98357A VIN

  GND

  GND

The motor pins are an example picked from the pins left free when no screen (LCD) or lamp is attached. Boards differ, so always check the pin names printed on yours, and avoid GPIO 0·3·45·46 (used at boot) and GPIO 35–37 (used by memory). Set the same numbers for the motor pins in the firmware.

1. Order the parts

   Order from the list above. Pick an ESP32-S3 camera board with 8MB PSRAM (N16R8). Overseas shipping can take a while, so buy spare wires at the same time.

2. Assemble the body

   Follow the two-wheel kit’s manual and screw the motors, wheels, caster wheel and battery holder to the plate. Push the board, microphone and amplifier into the breadboard and set it on the plate.

3. Wire it up

   Connect the wires following the pin map above. Battery + goes through the power switch to the motor driver (VM) and the voltage converter; the converter’s 5V goes to the board and the amplifier. Before switching on, check again that + and − are not swapped.

4. Flash the firmware

   Flash xiaozhi-esp32, an open-source voice AI firmware, to the board over USB. The easiest start is a ready-made file in a Chrome web flasher; to build it yourself, choose the breadboard board with a camera (bread-compact-wifi-s3cam) in ESP-IDF.

5. Connect Wi-Fi and the brain

   With your phone, join the Wi-Fi the robot creates (it starts with Xiaozhi-) and enter your home Wi-Fi (2.4GHz) at http://192.168.4.1. The robot reads out a six-digit code; register it in the xiaozhi.me console and choose a name, voice, language and AI model.

6. Add movement tools

   Copy the board folder to make your own robot board and register self.car.move (drive the wheels) and self.car.stop (stop) as MCP tools (code below). Now when you say “Go forward a little”, the AI calls that tool.

7. Test it

   Say the wake word or press the button on the board to start talking. Ask “What do you see in front of you?” and it takes a photo with the built-in camera tool (self.camera.take_photo) and answers. Test movement with the wheels lifted off the floor at first.

Whichever you choose, keep the AI’s API key on the server only, never in the robot. The robot only holds a device code.

1. xiaozhi.me — the easiest start

   The official xiaozhi server. Register the device code and start talking right away; individuals can use it for free (terms may change). It is run in China, though, so your voice and photos go there.

2. Your own server — pick any AI

   Run the open-source xiaozhi-esp32-server on your server with Docker and change the server address in the firmware. You choose speech recognition, the LLM (OpenAI, Gemini, DeepSeek, Ollama and more), speech synthesis and the vision model, and voice and photos pass through your own server.

3. Real-time voice AI — the most natural talk

   Connect straight to the OpenAI Realtime API, as ElatoAI does. One model hears, thinks and speaks, so answers are quick and natural (under 2 seconds round trip), but you pay for every minute of conversation.

The xiaozhi firmware has an MCP server inside the robot body. Register a tool’s name, description and inputs as below, and the AI reads that description during a conversation and calls the tool at the right moment. Drive() is a short function that switches the four motor pins above and stops the motors when the time is up.

// main/boards/my-robot-car/ (copy of bread-compact-wifi-s3cam)
auto& mcp = McpServer::GetInstance();

mcp.AddTool("self.car.move",
    "Drive the robot car. direction: forward, backward, left, right. ms: 100-3000",
    PropertyList({
        Property("direction", kPropertyTypeString),
        Property("ms", kPropertyTypeInteger, 100, 3000),
    }),
    [this](const PropertyList& p) -> ReturnValue {
        Drive(p["direction"].value<std::string>(), p["ms"].value<int>());
        return true;   // Drive() stops the motors when the time is up
    });

mcp.AddTool("self.car.stop", "Stop the motors now", PropertyList(),
    [this](const PropertyList&) -> ReturnValue { Stop(); return true; });

1. A walking robot like the video

   The Sesame robot in the reference video walks on four legs with an ESP32 board, eight MG90S servos and a small OLED screen for its face (about $50–60 in parts; needs a 3D printer and soldering). It has no microphone, camera or AI, so switch the board to an ESP32-S3 and add a microphone and amplifier to give it a brain the same way.

   Source Reference video

2. Raspberry Pi Zero 2 W

   Add a camera and the same microphone and amplifier to this $15 Linux computer and use AI APIs and OpenCV straight from Python (about $60–95 in total). It can do a lot, but it boots slowly and uses more power.

3. Start with a kit

   If choosing parts feels hard, start talking with a xiaozhi kit that bundles an ESP32-S3 camera board, microphone and speaker, such as the Keyestudio KS5028, then add wheels and a motor driver.