Robot Ross Skill

A Huenit robot arm that draws on paper while narrating in the warm, peaceful style of Bob Ross.

---

What Is This?

Robot Ross is an OpenClaw skill that controls a Huenit robot arm to draw/write on paper, narrate the process in Bob Ross style via a local AI model (Apertus 8B), and optionally engage visitors in live voice conversation.

Three modes:

• Headless — bob_ross.py is called directly (Telegram, scripts, order automation)
• Voice showcase — chat_ross.py listens via microphone, chats, and draws on request
• Pyrography — --pyro flag slows the feed rate for wood burning

---

Quick Start

Prerequisites

# Core (all platforms)
pip install pyserial anthropic

# Voice showcase
pip install openai-whisper sounddevice kokoro soundfile "numpy<2.0" espeakng-loader mistralai

# Image tracing (optional)
pip install vtracer pillow pillow-heif

Ollama must be running with Apertus loaded:

ollama pull MichelRosselli/apertus:8b-instruct-2509-q4_k_m

API keys

Keys are loaded from Infisical vault if available, otherwise from environment variables:

export ANTHROPIC_API_KEY=sk-ant-...   # for AI sketch generation + Haiku conversation
export MISTRAL_API_KEY=...            # for Voxtral TTS (optional)

Serial port

# Mac (default)
export HUENIT_PORT=/dev/cu.usbserial-310

# Windows
set HUENIT_PORT=COM3

---

Calibrate

python bob_ross.py calibrate

Press Q at any calibration step to abort and save Z height only.

---

Headless drawing (bob_ross.py)

# Write text
python bob_ross.py write "Hello World"

# Draw a shape
python bob_ross.py draw circle 40

# Draw an SVG file
python bob_ross.py svg path/to/file.svg --size 80

# AI-generated sketch
python bob_ross.py sketch "a lighthouse on a rocky coast" --size 80

# Pyrography (wood burning)
python bob_ross.py svg logo.svg --pyro --size 60

# Test narration without moving the arm
python bob_ross.py sketch "a happy little tree" --dry-run

# System check
python bob_ross.py check

Key flags:

Flag	Description
--size 80	Drawing size in mm (default 80)
--feed 400	Feed rate mm/min
--buyer "Alice"	Personalise narration with visitor name
--pyro	Pyrography mode — slow feed for wood burning
--engine kokoro|voxtral|system	TTS engine (default: kokoro)
--no-voice	Skip all narration
--dry-run	Narrate without moving the arm
--direct	Skip preview, draw immediately

---

Voice showcase (chat_ross.py)

A full voice conversation loop: visitor speaks → Whisper transcribes → brain (Haiku or Apertus) decides → Voxtral/Kokoro speaks → arm draws.

# Mac
python chat_ross.py --brain apertus --engine voxtral --lang en

# Windows
set HUENIT_PORT=COM3 && python chat_ross.py --brain apertus --engine voxtral --no-obs --lang en

# Test without arm movement
python chat_ross.py --no-arm --brain apertus --engine voxtral

Key flags:

Flag	Description
--brain haiku|apertus|mistral	Conversation LLM (default: haiku)
--engine kokoro|voxtral|system	TTS engine (default: voxtral)
--lang en	Force STT language
--whisper base	Whisper model size (tiny/base/small/medium)
--threshold 0.03	Mic RMS threshold (raise in noisy rooms)
--size 80	Drawing size mm
--no-arm	Conversation + narration only, no arm movement
--no-obs	Skip OBS recording

Press Q to quit gracefully, or say "goodbye" / "bye".

Conversation flow:

1. Robot Ross greets the visitor
2. Visitor describes what to draw
3. Robot Ross repeats the request and asks for confirmation ("Shall I go ahead?")
4. On "yes" → calibrates (first draw only) → draws while narrating
5. Conversation TTS mutes while arm is narrating — no audio overlap

---

Pre-made SVGs

Drop any .svg file in the repo root. chat_ross.py discovers them automatically and tells the brain about them so it can recommend them for instant draws (no AI generation wait).

---

Pyrography mode

--pyro reduces the feed rate to ~45 mm/min for slow, controlled burns on wood. Works with any drawing command:

python bob_ross.py svg design.svg --pyro --size 60
python chat_ross.py --pyro --engine voxtral

---

AI stack

Role	Model	Runtime
Showcase conversation	Claude Haiku 4.5	Anthropic API
Showcase conversation (local)	Apertus 8B	Ollama local
Bob Ross narration	Apertus 8B	Ollama local
AI sketch / SVG generation	Claude Haiku 4.5	Anthropic API
Speech-to-text	Whisper base	Local
TTS (showcase)	Voxtral mini	Mistral API
TTS (narration)	Kokoro 82M	Local
TTS (fallback)	System (PowerShell/say/espeak)	Built-in

Apertus is developed by MichelRosselli as part of an Innosuisse-funded Swiss AI project.

---

File structure

bobrossskill/
├── bob_ross.py             # Headless orchestrator (write/draw/svg/sketch/pyro)
├── chat_ross.py            # Voice showcase loop (STT → brain → TTS → draw)
├── ARCHITECTURE.md         # Full architecture diagram
├── huenit/
│   ├── huenit_svg.py       # SVG → G-code (Bezier-aware, path sort)
│   ├── huenit_write.py     # Text calligraphy (auto-scale, multi-line)
│   ├── huenit_draw.py      # Shapes + interactive calibration
│   ├── huenit_jog_control.py  # Manual jog utility
│   ├── huenit_wave.py      # Greeting gesture
│   └── calibration.json    # Saved Z/tilt calibration (auto-generated)
└── voice/
    ├── speak.py            # TTS: Kokoro / Voxtral / system fallback
    └── listen.py           # STT: Whisper with VAD silence detection