A suitcase robot changes its speech when a gas sensor detects smoke

Sparky is an offline suitcase robot that uses an to detect smoke and certain airborne chemicals. Every 0.5 seconds, it compares the current sensor reading with a changing clean-air baseline. The signal becomes a 0 to 10 phase: it rises when smoke is blown at the robot and slowly falls over several minutes. That phase changes the LLM sampler while each answer is being generated.

moves from 1.0 to about 1.6, top_p from 0.95 to 0.99, and top_k from 64 to 120. As the phase rises, the robot chooses less predictable words and sounds more scattered. The behavior is not a fixed script, so each response is newly generated and does not simply repeat. Small persona hints, a slower voice, drooping and bloodshot eyes, and a sensor display also change with the phase.

At phase 10, the display shifts into an intense smoke-and-plasma effect and keeps the robot in that state for about 7 minutes. The sensor can also react to cigarettes, incense, or other vapor-like smells, not only one specific kind of smoke.

Key points

  • Sparky runs offline inside a suitcase-style robot.
  • An checks smoke and vapor-like signals every 0.5 seconds.
  • The sensor reading becomes a 0 to 10 phase that rises and then slowly fades.
  • Higher phases change the LLM sampler, making replies less predictable.
  • The robot also changes voice, eyes, and screen effects to match the phase.
Read original