A working cat, not a startled puppy

By the afternoon of 20 October the robot could hear, turn towards a sound, play a small emotion and find my face. So I let it live on the desk for an hour while I worked.

The test

From 16:50 to 17:46 it reacted 291 times. One reaction every eleven or twelve seconds, for an hour.

What it did	How many times
Startled, from a sound 12 dB above the background	105
Turned towards what it took for a voice	46
Emotions used, out of the 85 available	4

My three notes at 18:00: it reacts to everything; it does not actually look at the noise; its emotions look random and always the same.

The principle

Before touching any number, I wrote down what I wanted, in one image: a working cat. Present, a bit funny, it gets used to the sounds of the house, it glances at what is new, and it ignores what it knows. Not a puppy that jumps at every door.

Inside the machine

A microphone does not hear “a door”. It measures a level. The program compares each short window of sound with the background noise of the room, in decibels above that background, and sorts it:

Level above the background	Taken as
from +8 dB	speech
from +16 dB	a knock
from +26 dB	a loud knock

Then three mechanisms, all borrowed from animals:

1. Rest. After a startle, 8 seconds without startling again; after a turn, 15 seconds.
2. Habituation. Each reaction raises the bar by 2 dB for 30 seconds. A noise that repeats ends up ignored.
3. A glance first. The head turns a little towards the noise, at most 35 degrees, before any emotion. It looks, then it reacts.

And one rule for my voice. I dictate a lot, and dictation at 50 centimetres reaches +43 dB. So my voice never makes it startle, except a shout. It only answers a sentence addressed to it: between 0.4 and 4 seconds long, after 2 seconds of silence, with my face turned towards it.

What the numbers said

Evening period	Result
18:11 to 18:17	one reaction every 11 s; my dictation still dominated
18:28 to 18:53	one reaction every 35 s, and 43 dictated sentences ignored

From every 11 seconds to every 35. Better. Not good.

What still broke

Five days later, living with it, the verdict was the same in fewer words: far too active for nothing. A slightly raised voice was enough to trigger an animation. In the night of 24 to 25 October I raised every threshold by 6 dB, set the rest after a reaction to 60 seconds, and removed the strong emotions altogether.

The real fix is not a number, it is a sense. A sound alone should rarely be enough: the robot should react when someone actually comes close, within a metre and a half or two, which means crossing what it hears with what it sees. Before writing those thresholds, I want them measured in decibels on my own desk, not guessed.

Next: the robot was saturated, and not by what I thought.