BMO

A voice-activated robot friend on a Raspberry Pi 5, inspired by Adventure Time. Built from scratch by Corey and Sylas (age 11) as a learning project covering AI, speech, Linux, and game emulation. Full design doc: bmo_plan.md.

BMO boots straight into a fullscreen animated face and acts as the whole OS:

• Say "BMO" (or touch the screen) to wake it, then just talk — a local LLM answers out loud. No cloud, works with zero internet.
• Launches NES, SNES, Sega, and Game Boy games by voice: "BMO, play Super Mario World" starts RetroArch fullscreen; quitting the game (Esc or the pad hotkey — the mic is off while a game runs) brings the face back.
• Timers, alarms, and a stopwatch; jokes, riddles, stories, and word games; music, weather, volume control.
• Long-term memory: "BMO, remember that my favorite color is teal" — facts survive reboots and the LLM knows them in every conversation.
• Kid-safety guardrails: bad words, grown-up topics, and anything worrying get a gentle scripted redirect (never the LLM), on both what BMO hears and what it says. Matches are logged for parent review.
• Conversations stay open ~2 minutes after the last exchange; "BMO, stop" puts it back to sleep.

How it works

Piece	Tech	File
Face	pygame-ce, parametric (no images)	bmo/face.py
Hearing + wake word	Vosk small model, grammar-restricted while sleeping	bmo/ears.py
Voice + lip sync	Piper TTS → sox pitch shift → RMS envelope drives the mouth	bmo/voice.py
Brain	Ollama: qwen2.5:3b (default) or llama3.2:1b, streamed sentence-by-sentence	bmo/brain.py
Commands	regex intent router; plugins win, LLM is the fallback	bmo/router.py, bmo/plugins/
Memory	diary facts folded into the system prompt	bmo/plugins/diary.py, var/diary.json
Safety	3 tiers (crisis / profanity / grown-up), input + output	bmo/plugins/safety.py

Everything is configured in config.toml.

Things to say

After "BMO" (or a touch), try:

• Games — "play Super Mario World", "play Aladdin on Genesis", "what games do we have", "what Mega Man games do we have", "do we have Sonic"
• Videos — "watch minecraft parkour", "play the mario theme on youtube": BMO shows the top 6 hits; pick by D-pad + Start, tap, or voice ("the first one"). Say "BMO" (or press Start) to stop the video.
• Time — "set a timer for five minutes", "set an alarm for 7 30 a m every school day" (also "every day"; up to 3 alarms, they survive restarts and ring for 3 minutes), "cancel my alarms", "start the stopwatch", "how long has it been", "what time is it"
• Fun — "tell me a joke", "tell me a riddle", "tell me a story", "let's play twenty questions", "play a word game"
• Music & weather — "play some music", "play the song …", "next song", "what's the weather"
• Memory — "remember that …", "what do you remember", "forget …"
• Robot stuff — "louder" / "quieter" / "set the volume to 50", "use your little brain" (fast model) / "big brain" (smart model), "what can you do", "stop", "good night"

Anything else goes to the LLM — just talk.

Running on the Pi

BMO runs as a systemd user service: it autostarts on boot, restarts itself after a crash, and appears in the Start Menu as BMO. Manual control:

systemctl --user start bmo      # or stop / restart / status

• Touch anywhere: wake BMO / interrupt it mid-sentence.
• Tap the mic button (bottom-right circle): mute/unmute the microphone. Red with a slash = BMO can't hear anything until you tap it again — touch and typing still work.
• Type + Enter (any keyboard): chat without voice — also works before a mic is attached.
• Hold the top-right corner for 2.5 seconds (or press Escape), or say "exit to desktop": exit-to-desktop dialog for grown-ups — the screen washes red left-to-right while you hold, no keyboard needed. Exiting always takes a touch on the dialog, so a misheard phrase can't close BMO by itself.
• Drop game ROMs in roms/nes/, roms/snes/, roms/genesis/, roms/sms/, roms/gamegear/, roms/gb/, or roms/gbc/; songs in music/.
• Running timers, set alarms, and the stopwatch show as small pills in the top-left (live countdown). Tap one to adjust (+1 min), cancel, or close; when something rings, its pill flashes red — tap to hush it.
• The controller works on BMO's face too: Start wakes BMO or interrupts it mid-sentence. An ambiguous "play zelda" opens an on-screen picker — D-pad to choose, Start/A to launch, B/Select to cancel (tapping a row or saying the full name also works).
• In-game controller combos (hold Select plus): Start quit back to BMO, R save state, L load state.
• In-game keys: F2 save state, F4 load state, hold R to rewind time, Esc quit back to BMO.
• Conversation transcripts land in logs/ as JSONL; safety triggers are logged there with "kind": "safety". BMO's diary is plain JSON in var/diary.json — easy to review or edit.

Fresh Pi setup from a clone:

sudo git clone https://github.com/coreyms/bmo /opt/bmo
cd /opt/bmo && ./install/setup.sh

The setup script installs apt packages (RetroArch + NES/SNES/Genesis/Game Boy cores, mpv, sox, portaudio), Ollama + both models, the Vosk and Piper models, the systemd user service, and the RetroArch quit hotkeys (Esc or Select+Start).

Power matters: a Pi 5 needs a real 5V/5A (27W) USB-C supply. Phone chargers negotiate 15W and cause brownout crashes under LLM load. Ask us how we know.

Dev setup on Mac or Ubuntu

BMO runs windowed on a laptop — same code, --dev flag. Games/autostart are Pi-only; everything else (face, chat, TTS, STT, plugins) works anywhere.

1. System packages

# macOS (Homebrew)              # Ubuntu/Debian
brew install sox mpv            sudo apt install sox mpv libportaudio2 python3-venv

(macOS doesn't need portaudio — the sounddevice wheel bundles it.)

2. Ollama + models — install from ollama.com (or brew install ollama), then:

ollama pull qwen2.5:3b
ollama pull llama3.2:1b

3. Python env (needs Python 3.11+):

git clone https://github.com/coreyms/bmo && cd bmo
python3 -m venv .venv
.venv/bin/pip install -r requirements.txt

4. Speech models (~100 MB, into models/):

curl -L -o /tmp/vosk.zip https://alphacephei.com/vosk/models/vosk-model-small-en-us-0.15.zip
unzip /tmp/vosk.zip -d models/
mkdir -p models/piper && cd models/piper
../../.venv/bin/python -m piper.download_voices en_US-amy-medium
cd ../..

5. Run it:

.venv/bin/python -m bmo.main --dev

Click to wake BMO; type + Enter to chat. With a mic and speakers attached, voice works exactly like on the Pi. Escape opens the quit dialog.

Writing a plugin

Add a file in bmo/plugins/ — it's auto-discovered. Minimal example:

from bmo.router import Plugin, Result

class HighFivePlugin(Plugin):
    name = "highfive"
    priority = 45          # lower runs first; LLM handles anything unmatched

    def __init__(self, app):
        super().__init__(app)
        self.add(r"\bhigh five\b", self.slap)

    def slap(self, match, text):
        return Result(speech="Up top! Smack! That was a good one.")

Return Result(speech=...) to answer directly, or Result(brain_text=..., style_hint=...) to hand the request to the LLM with steering. Plugins can reach everything through self.app (voice, brain, timers, config) and may define a tick() method that runs every frame.

Project layout

bmo/            the program (face, ears, voice, brain, router, main)
bmo/plugins/    one file per ability — add yours here!
install/        setup script, supervisor, desktop entry
config.toml     models, voices, wake phrases, timeouts
bmo_plan.md     the full design document (worth reading!)
roms/ music/    your games and songs (gitignored)
models/ logs/   downloaded speech models, transcripts (gitignored)

License / credit

Personal, non-commercial fan project. BMO and Adventure Time belong to Cartoon Network; this is a kid's learning project, not a product.