How to Build a Shitty Robot

2 days ago

A low-cost toy robot was purchased and disassembled to create an LLM-powered, voice-controlled robot for children, emphasizing simplicity and readily available materials.
The robot's original PCB was reverse-engineered, revealing a simple design with a single motor using gears for walking and turning; the main IC was removed to hijack the H-bridge for motor control.
A new chassis was built from cardboard and a paper cup, housing the electronics and a phone, which serves as the display and audio interface, controlled via an Adafruit FT232H board for GPIO.
A client-server architecture was implemented, with the server handling speech-to-text, text-to-speech, LLM, and agent logic locally for privacy, while the client runs in a browser using WebUSB (except Safari).
Speech-to-text uses Parakeet TDT 0.6B with Silero VAD for real-time transcription, and text-to-speech uses Qwen3 TTS for voice cloning, both optimized in Rust to avoid Python dependencies.
The pipeline supports barge-in for interruptions, low-latency responses via sentence chunking, and tools like web search, photo capture, motor control, and Spotify integration, powered by local LLMs like Gemma 4 26B.

Hasty Briefsbeta