Skip to content
← The Workbook

No. 56 · 2026Concept

Corvid

A computer-vision crow feeder that trains the local birds to bring me things. Camera identifies crows in the yard, reads their behavior, and dispenses tiered food in response. Higher-value treats for higher-value deliveries. Long-term: the birds learn that bringing shiny objects, lost keys, or interesting curiosities to a specific tray gets them the good food.

Status
Concept
Year
2026
Stack
Computer Vision · Animal Training · Outdoor Hardware · Pi · Jetson

The Idea

Crows are smart enough to learn an exchange relationship with a feeder. Joshua Klein proved this in his 2008 TED talk with a crow vending machine that trained urban birds to deposit coins. Corvid is the at-home version aimed at a more interesting trade: bring me something cool, get the good treats.

Camera in the yard watches for crows. Computer vision identifies them (corvid silhouette + behavior pattern, not just any bird at the feeder), reads what they're doing (just visiting, dropping something off, hanging around the deposit tray), and the dispenser responds accordingly.

The Tiered Reward

  • Visiting tier. Crow shows up, gets recognized, gets a small everyday treat. Builds the association: this feeder is for me, this is a safe stop on the route.
  • Pattern tier. Crow comes back regularly, gets a slightly better treat. Reinforces the feeder as a reliable stop.
  • Delivery tier. Crow drops something on the deposit tray. The good stuff, the high-protein, high-novelty treat the crow will remember. The delivery doesn't have to be valuable to me; the bird just has to be doing the thing.
  • Repeat-delivery tier. Crows that consistently deliver get the best food. Conditioning compounds. The most reliable birds become the most trained.

The point is not to extract objects from the birds (that's a side effect). The point is to engineer a long-running exchange relationship between the household and a small group of intelligent urban animals.

Hardware Concept

  • Camera node. A small Pi or Jetson with a corvid-tuned vision model. Local inference, no cloud.
  • Dispenser. Multi-hopper, servo-driven, holds 3-4 tiers of treats. The same hardware family as the Fetch dog dispenser.
  • Deposit tray. A flat surface in front of the camera where the bird can drop things. A second camera or weight sensor detects the drop.
  • Optional: a small claw / drawer that retrieves the deposited object to a covered bin so the next visitor can't grab it back.

Pairs With

Same instinct as the dog cluster. Fetch (CV trainer for dogs), Drover (working dogs for human herding), Barkangel (the dog GPS that the dog charges), applied across species. Animals as legitimate users of household interfaces, given the right hardware and the right incentives.

What Could Go Wrong

Honest list:

  • Habituation. Feeding wild animals creates dependence and changes their territory range. Need to dose carefully and consider the long-term effect on the local population.
  • Neighbors. A reliable food source brings more crows than just yours. Some neighbors will not love this. Worth talking to people on the street before scaling it up.
  • Disease vector. Shared feeding surfaces are how avian diseases spread. The dispenser hardware and deposit tray need cleaning protocols, otherwise the project actively harms the birds it's trying to support.
  • Reading "behavior" with CV. Crow recognition is doable (silhouette, posture, common behaviors are well-studied). Distinguishing "just visiting" from "depositing" needs careful sensing, easy to misclassify and reward the wrong thing.
  • Manipulation framing. Training wildlife with a tiered conditioning system has an ethical edge. The honest version says so. The fun version doesn't pretend otherwise.

Status

Concept. The hardware is straightforward (Pi or Jetson + camera + multi-hopper servo dispenser + deposit-tray sensor). The interesting work is the corvid vision model and the deposit-recognition logic. Joshua Klein's crow vending machine is the prior art proving the bird side actually works.