How it works
- Watch. The robot camera streams about one frame per second to
gemini-robotics-er-2-streaming-previewthrough the Gemini Live API, together with your microphone audio. - Reason. The model answers in short spoken sentences (Gemini TTS on the robot speaker) and decides which body tools to call.
- Look. Tools like
look_ataskgemini-robotics-er-2-previewto point at the object; the SDK converts the pixel into a head pose. - Express. Nods, dances and emotions come from Pollen's recorded move libraries; face and object tracking keep the head on target.
Two modes
🎙 Voice
Live bidirectional session: talk naturally, interrupt, ask "what am I holding?", "look at the cup", "tell me when I raise my hand".
⌨️ Text
Turn-based chat in the web page: each message ships a fresh camera frame, tools run in a loop, replies can still be spoken.
Tools the model can call
look_at·find_objects·track_object·follow_facemove_head·play_emotion(40 recorded moves) ·watch_for·describe_view
Setup
- Install from the Reachy Mini dashboard and start the app.
- Open the app page (⚙️ icon in the dashboard,
http://localhost:8042on a Lite). - Paste a Gemini API key (free tier at aistudio.google.com/apikey).
- Pick Voice or Text and say hi.