2023.07.20

Building an AI Plotter Bot.

Turning a CNC plotter into a GPT-4-powered drawing machine that turns text prompts into SVG line art.

GPT-4CNCRoboticsSVG

A CNC plotter draws by moving a pen on two axes. I had a small belt-driven plotter collecting dust. The idea: connect it to GPT-4 vision to convert text prompts → SVG → G-code → physical drawing.

Components

2-axis belt plotter (CNC shield + GRBL firmware)
Raspberry Pi 4 as the controller
GPT-4 Turbo for prompt-to-SVG generation
Node.js pipeline script
svgToGcode library for path conversion

Prompt to SVG

const { OpenAI } = require('openai');
const client = new OpenAI();

async function promptToSVG(prompt) {
  const completion = await client.chat.completions.create({
    model: 'gpt-4-turbo',
    messages: [{
      role: 'user',
      content: `Generate a clean SVG line drawing of: ${prompt}.`,
    }],
  });
  return completion.choices[0].message.content
    .match(/<svg[\s\S]*?<\/svg>/)?.[0];
}

Send G-code to plotter

# Convert SVG to G-code
node svg2gcode.js output.svg > plot.gcode

# Stream to GRBL over USB serial
node stream.js plot.gcode /dev/ttyUSB0

GPT-4 Turbo is inconsistent at generating valid SVG. About 30% of outputs had syntax errors or coordinates outside the viewbox. I added a validation pass with an XML parser and a simple bounds check.

What worked well

Simple geometric prompts: 'mountain silhouette', 'face outline', 'leaf pattern'
Abstract geometric art: reliable and clean
Portraits: hit-or-miss but occasionally impressive

The full pipeline from text prompt to physical drawing takes about 90 seconds. Slow, but the output is physical and that makes it feel like magic.

↳ Use a felt-tip pen, not ballpoint — the plotter lacks enough Z-axis pressure for reliable ballpoint contact.