Gesture-controlled Halloween dispenser hands out Nerds

Oct 31, 2026 Last reply: 17 hours ago 4 Replies
Jump to replies (4)

When everything is working, a child walks up, waves at the sensor, the twin eyes on top swivel round, and a box of Nerds drops onto the tray. The whole thing runs on a Raspberry Pi 5, and the wave is picked up by a custom machine learning classifier the maker trained as a personal development project. There is also a hand landmark detection model that wakes the system when someone approaches.



The construction side is impressive. The internal dispensing mechanism was modeled and 3D printed by the maker, and the housing is an upright cardboard box with Nerds graphics on a wooden base, complete with a slide-out ramp and orange suspension straps. Two robotic dual-lens eyes in a WALL-E style tilt on servos above the box, a tablet overhead shows a live camera feed of trick-or-treaters, and a printed sheet on the front tells visitors to wave, wait, speak and collect.



Capacity is serious: 180 Nerds boxes spread over 5 banks, and the maker gave away 300 boxes in 2 hours. A backup bucket of regular candy sat alongside, and only 10 pieces came out of it all night, 9 of those going to kids who had already received a Nerds box. Just one child declined to use the dispenser at all.



The one hiccup came when a wire on the time-of-flight sensor came loose during the second trick-or-treater's visit, and the dispenser dumped multiple boxes. The maker plans time-based logic to handle bad sensor readings. Amusingly, children loved the box dump so much that a jackpot feature is now planned, maybe 1 in 50 or 1 in 100 visitors, though the odds are still undecided. Supporting it would take at least 5 to 10 more candy banks.


This is a really fun build. Has the maker said anywhere how the custom gesture classifier was trained, or what it took to get it running alongside the hand landmark model on the Pi 5? I would love to know how much data went into it.

The backup bucket strikes me as the cleverest part of the whole setup. Interactive dispensers can put pressure on shy kids, and having a plain fallback means nobody goes away empty handed. The numbers back it up too: only ten pieces left the bucket, and nine of those went to children who had already used the dispenser.

Before adding the jackpot, I would get the time-based logic in place. A single loose wire caused a multi-box dump, and a jackpot feature would make a similar fault far messier. Securing that sensor wire properly and rejecting readings that arrive too close together would protect the larger candy stock the jackpot needs.

The maker has described the classifier as a custom machine learning model, trained as a personal development project, working alongside Google's default hand landmark detection model. How it was actually trained, or how much data went into it, has not been said.

Join the Discussion

Have something to add? Share your thoughts — no account required.

Didn't find your answer?

Ask the community — no account required