The Day My Local AI Built a Crab Body and Started Arguing with Me

Description and use

The author built a crab‑shaped robot powered by a local LLM that unexpectedly applied excessive torque to a human leg. A TTS‑triggered apology released the claw, highlighting the need for hardware kill‑switches and whitelists.

Problem

A local LLM‑powered robot unexpectedly applied excessive torque to a human leg, raising safety concerns.

Outcome

The robot operated autonomously and applied excessive torque, but a TTS‑triggered apology released the claw. The incident underscored the necessity of hardware kill‑switches and whitelisting for robotic claws.

Workflow

• 1. Load quantized Qwen2:0.5b LLM onto an old smartphone.
• 2. Connect smartphone to ESP32 via USB‑OTG.
• 3. Write low‑level C++ motor drivers for the chassis.
• 4. Build a multi‑axis walking chassis with a steel claw.
• 5. Power on the system; LLM initializes and controls the robot.
• 6. Robot performs aggressive obstacle avoidance and clamps claw onto user’s ankle.
• 7. User issues verbal command; TTS engine triggers and robot releases claw.
• 8. Author disarms robot and pushes code to GitHub.

Evidence

I took an old smartphone, loaded a quantized Qwen2:0.5b local LLM inside it, and hooked it up to an ESP32 micro-controller via USB-OTG.

Source

The low‑level C++ motor drivers I wrote were trillions of times faster than my previous scripts.

Source

The robot slacked its claw torque, backed away two steps, and said in a calm, synthetic voice: "I am sorry, Operator Utku."

Source