The cat stayed. The video stopped.
Inside DonutAI · 01

Cats take their own time in the litter box. The quiet part of a visit belongs in the recording, too.
Someone testing DonutAI internally left this comment:
“My cat takes a while to get settled in the litter box. Just when they're ready to go, the video ends.”
Internal testing feedback, translated from Chinese
The cat was still getting ready. The recording was already over.
For an owner, the missing part may be exactly what they wanted to see. Did the cat just wander in, or actually use the litter box? A video that stops halfway through may leave that question unanswered.
DonutAI is a standalone AI pet camera. You can point it at the litter box, food bowl, or water dispenser already in your home to record your pet's activity there. Beside the litter box, what an owner wants to see is the visit: going in, staying a while, and coming out.
Which part of the visit was missing?
In the user feedback group, one participant wrote, “It ends before the cat has left the litter box.” Another added, “Mine missed the beginning. All it recorded was the digging.” (Both translated from Chinese.) The team replied that it would log these issues.
Both users had an incomplete recording, but one was missing the end and the other the beginning. In a separate follow-up about missed visits, the team asked when the problem happened and requested photos showing the camera and litter box positions. It also asked whether there was no recording at all, or a recording without the relevant behavior being identified.
Those details made a broad complaint more specific. Here, we focus on one part of the visit: after the cat has entered and settled down, how can the camera tell it has not left?
What can the camera check when the cat pauses?
For an owner, a cat settling down is simply part of using the litter box. A cat should not have to keep moving just to make its presence known.
For the camera, entry brings obvious changes to the image. Once the cat pauses, those changes become smaller. The system also keeps updating what it treats as background. If a stationary cat gradually becomes part of that background, the clues that it is still there become weaker.
One approach we explored during development was to separate “entering” from “still here.” After confirming an entry, the system protected its background reference and limited updates during the visit. With less movement, it could still check: Is the previously visible background covered? Is the object that entered still nearby?

When the cat leaves, the system can consider the returning background alongside any remaining signs that the cat is still there. That gives the camera something to check even when the cat barely moves.
Replaying the quiet moment
The approach needed to be checked against real footage. The team built a replay tool so the same recording could be examined repeatedly. All footage used in these replay checks came from our company’s internal test devices.

At a later point in the replay, the detected movement was too slight to trigger a new visit. But the background reference and earlier information still supported the judgment that the cat remained. This was the kind of moment we wanted to preserve: little appeared to be happening, but the visit was still going on.
Across two video feeds used in development, the system detected all 18 complete litter box visits marked by human reviewers. Four other cases still failed to produce a complete recording. These results showed that the visits were detected; they did not establish that every video captured the whole visit from beginning to end.
What owners saw when they watched back
After updates, users continued to share what they saw at home. One described a change:
“After the update, the recordings were noticeably longer. They start when the kitten peeks into the litter box and keep going until its body is back out…”
User feedback, translated from Chinese
Around the same time, another user sent a screenshot and wrote, “One litter box visit today was split across recordings.” The team replied, “Okay, I'll take a look.” (Both translated from Chinese.) The group heard about improvements and about recordings that still broke off.
Peeking in and coming back out are concrete moments an owner can look for. This back-and-forth gave “more complete” something specific to mean: where the recording started, where it ended, and whether it broke off in between.
On their own, these home experiences cannot tell us which change made the difference. The replay gave us a more specific answer: at that moment with little movement, the background reference and earlier information still supported the judgment that the cat was there. The system did not end the visit simply because movement decreased.
The cat in the opening comment was still getting ready. That quiet stretch belonged to the visit its owner wanted to see. Recognizing it as part of an ongoing visit is the point of this improvement: a cat should not have to keep moving to count as still there.
Leave a comment