空間 AI エージェントを構築する — デプロイガイド

1. Confirm vision detection mode: say to Watcher "Enable face detection" or "Set vision mode to 1" (self.vision.mode: 1). Watcher confirms entering on-device person/face detection mode.
2. Place the configured Watcher on your desk at face height (distance ~0.5 to 1.5 m).
3. Walk up to Watcher and pause for longer than the validation duration (default 3 s):
   • When the camera detects a person present, Watcher switches expressions and greets you proactively (e.g. "Hello! Good to see you!"), entering conversation without requiring a wake word.

During the conversation, ask Watcher about its vision settings:

1. Connect Watcher and select the SERIAL-A port (the lower-numbered port on macOS/Linux, e.g. ttyACM0).
2. Flash Himax WE2 AI models (SCRFD 500M + MobileFaceNet 128D).

Face database embeddings are stored in ESP32-S3 local NVS flash:

1. Select the SERIAL-B port (the ESP32-S3 serial port).
2. If already flashed in stage 1, select the port to connect; if not, flash the Xiaozhi firmware.

Connect via USB to manage the local face database:

1. Click Connect to start real-time camera preview (reads from Himax port).
2. Click Enrol in the Face Database panel (reads/writes NVS via ESP32 port) and enter a name.
3. Face the camera in good lighting for ~5 seconds to extract the 128D embedding.
4. Up to 20 faces stored in on-device NVS.

Disconnect USB cable; Watcher runs standalone:

1. Enable Recognition Mode: Say to Watcher "Enable face recognition" or "Set vision mode to 2" (self.vision.mode: 2).
2. Familiar Greeting: When an enrolled person approaches, Watcher extracts features locally and greets them by name (e.g. "Hello, John!").
3. Stranger Alert: When an unrecognised person approaches, Watcher notes an unknown visitor.
4. Voice Commands:
   • "Who do you know?": lists enrolled names.
   • "Can you recognise me?": captures live frame and tests recognition.
   • "Enable familiar mode": mutes greetings for known faces (mode: 3); only alerts on strangers.