docs: document on-robot demo-station setup and tuning

README.md

@@ -85,10 +85,36 @@ python -m mr_reachy
 Say "what do you see?" to trigger the vision path (once a vision sub-account is funded).
 Say "bye" / type `quit` to exit.
 
-## On real hardware
+## On the robot (demo station — Reachy Mini Wireless)
 
-Connect to the robot's daemon instead of the sim and run the same loop, or
-publish to the Reachy Mini app store:
+At the demo station the app runs **on the robot's Pi**, started from the dashboard
+for a short slot. On hardware Mr Reachy uses the robot's **onboard mic, speaker
+and camera** (via the daemon media manager) and **antenna push-to-talk** instead
+of the laptop mic — selected automatically by `RobotBackend` when published, or
+with `--on-robot` when driving a robot from your machine.
+
+> ⚠️ The on-robot I/O path is built against the SDK API but has **not been tested
+> on a physical robot yet**. Rehearse before the slot and tune the two values below.
+
+**1. Speech — install Piper on the robot** (0G has no TTS provider):
+
+```bash
+pip install "mr-reachy[robot]"          # piper-tts + pillow
+# download a voice once, then point the app at it:
+export PIPER_MODEL=/path/to/en_US-amy-medium.onnx   # (+ matching .onnx.json)
+```
+
+**2. Internet** — the robot must reach the 0G endpoints
+(`compute-network-*.integratenetwork.work`). Venue Wi-Fi may block this; test early.
+
+**3. Antenna push-to-talk** — press an antenna to talk, release to send. Tune the
+sensitivity if needed:
+
+```bash
+export MR_ANTENNA_THRESHOLD=0.5         # radians of deflection that counts as a press
+```
+
+**4. Publish & install:**
 
 ```bash
 # Run the checker DEACTIVATED — an active venv's $VIRTUAL_ENV breaks its nested temp venv:
@@ -96,17 +122,32 @@ env -u VIRTUAL_ENV .venv/bin/reachy-mini-app-assistant check .
 reachy-mini-app-assistant publish        # requires `hf auth login`
 ```
 
+Then **Install to Robot** from your Hugging Face Space (dashboard URL
+`http://reachy-mini.local:8000`) and start it from the dashboard. The robot
+auto-returns to its rest pose when the app stops.
+
 The app is registered via the `reachy_mini_apps` entry point
 (`mr_reachy.main:MrReachy`) in `pyproject.toml`.
 
+### Driving a robot from your machine (Lite, or testing Wireless over the network)
+
+```bash
+python -m mr_reachy --on-robot          # uses robot media + antenna push-to-talk
+```
+
+`ReachyMini()` auto-connects to a local daemon, falling back to `reachy-mini.local`.
+For an explicit network connection use `connection_mode="network"`.
+
 ## Layout
 
 ```
 mr_reachy/
   main.py        # ReachyMiniApp + conversation loop + CLI
   og_client.py   # 0G Compute: chat / STT / vision (OpenAI-compatible)
+  io_backends.py # Local (laptop) vs Robot (onboard media + antenna push-to-talk)
+  tts.py         # on-device speech synthesis (Piper / say / espeak)
   expressions.py # emotion -> head pose + antenna gestures
-  audio.py       # mic VAD capture + local TTS
+  audio.py       # laptop mic VAD capture + local TTS
   config.py      # loads .env into typed settings
 ```
 

agents.local.md

@@ -17,7 +17,15 @@
   reachy-mini-daemon --sim --headless --no-media --no-preload-datasets
   ```
   Use `--no-headless` (drop `--headless`) to open the MuJoCo viewer for demos.
-- Audio is handled **in-app** (laptop mic via sounddevice + macOS `say`), not via the daemon.
+- Audio is handled **in-app** via a backend (`io_backends.py`):
+  - **LocalBackend** (sim/dev): laptop mic (sounddevice VAD) + macOS `say`/espeak.
+  - **RobotBackend** (demo, Wireless): robot onboard mic/speaker/camera via `ReachyMini.media`
+    (`start_recording`/`get_audio_sample`, `play_sound`, `get_frame`) + **antenna push-to-talk**.
+    Speech is synthesized on-device with **Piper** (`tts.py`); set `PIPER_MODEL`. 0G has no TTS.
+- **Demo station = Reachy Mini Wireless**, app runs ON the Pi (Install to Robot from HF Space).
+  Robot needs internet to reach `compute-network-*.integratenetwork.work`. Tune
+  `MR_ANTENNA_THRESHOLD` (rad). RobotBackend is **not yet tested on physical hardware** —
+  unknowns: antenna press threshold, whether `media.play_sound` blocks, mic samplerate.
 
 ## 0G Compute (mainnet, chain 16661)
 - Secrets live in `.env` (gitignored). Three services, each its own host + `app-sk-` key.