Want to learn more? Check out our official website, where you can download a kit design and help develop the process!!
Below is the original page for the project.
The goal of this project is to create a device that will autonomously train crows. So far we’ve trained captive crows to deposit dropped coins they find on the ground in exchange for peanuts. The next step is to see how quickly we can get wild crows to learn the system, and then how quickly they can learn it from each other.
Once we’ve got system down for teaching coin collection we’ll move to seeing how flexibly they can learn *other* tasks, like collecting garbage, sorting through discarded electronics, or maybe even search and rescue. The crows continue to amaze us with their abilities, so who knows?
The first version of the device consists of a box from which protrudes a perch, a food tray, and a funnel. The whole thing is made out of sealed wood so as to minimize noisy clanging which might result from using metal components while retaining the ability to leave the thing out in the rain. It is run by a laptop which provides power and control up to 50 feet away.
Based on established Skinnerian training principles, the action of the device is divided into four stages:
Stage One: Food and Coins Available on Departure.
- At this stage the device pushes a few peanuts and one or two coins onto the feeder tray whenever a crow *leaves* the device. This ensures that the device always has food whenever it is examined by a potential feeding crow. It also ensures that both the sound of the device and its mechanical operation occur in close proximity to the feeding act so as to aclimate the crow. By having this noise occur as the crow leaves it prevents startling a potential feeder away from using the device.
Stage Two: Food and Coins Available On Landing.
- Herein the action of the device is identical except that food and coins are issued when a crow arrives. At this point the crow should be comfortable with the sound of the device and is now being trained to wait for its reward when arriving at the machine. Note that the feeding tray is slanted such that coins will pile up and prevent peanuts from being available until the crow cleans them away – a typical behavior of crows is to sweep things out of the way with their beak, and in this case this causes the coins to fall down the funnel. This should help reinforce the connection between coins going down the funnel and peanuts being produced.
Stage Three: Coins Available On Landing, Food Available on Deposit
- This is the highest-risk segment of the machine’s operation. At this point coins alone are made available whenever the bird lands on the perch. However, should a bird peck or sweep coins off the tray and cause a coin to fall down the funnel, the device then produces some peanuts. This stage is designed to cement in the crows’ mind the relationship between coins going down the funnel and peanuts being made available.
Stage Four: Food Available On Coin Deposit
Finally we shift the device into its intended, and long-term state of only providing peanuts when coins go down the funnel. Nothing is otherwise provided aside from coins scattered around the device at the beginning of the project.
If you’d like more details, I wrote a paper on the vending machine.
Press & Media
Previously I have presented on this topic at TED, Gadgetoff, and The Swedish Computer Science Institute. I also gave a 15-minute presentation on it as part of ITP’s thesis week. If that doesn’t grab you, Gizmodo did a nice little video review of it.
Articles have been written in Oprah Magazine, The Seattle PI, BoingBoing.net, Spice Magazine, The New York Times once, and then again in their Top Ideas of 2008 (which was followed by a “correction“), Jeff Jarvis’ BuzzMachine, and Wired.
Oh, and if you speak German, it was also featured in Vögel Magazine.
We’ve made a Creative Commons licensed, freely available design for the crowbox that anybody can download or buy as a kit. That means that anyone anywhere can make a version of the box and share their results with others, and we can test the machine with wild crow populations to see how to optimize the training process.
You can order it here. Alternately, feel free to donate to help us do more R&D on improving the design and conduct research into similar programs for other species.
Also see Zach Eveland of Blacklabel Development for more mad genius – he’s responsible for the electronics that made the original version of this project work.