I don’t want to tell these baby chickens how to live, but they’re going about their business all wrong. The cylindrical robot in their pen looks nothing like a hen, and it makes decidedly un-hen-like beeps, yet the chicks trail it obsessively, as if it’s their mother. Where the PoulBot goes, so too go the yellow little fluffs. Beep beep beep, says the robot. Chirp chirp chirp, say the chicks.
The idea behind this pairing, developed by researchers from several European universities, isn’t to give the chicks a complex—I promise—but to parse the extreme complexities of animal behaviors, especially as those behaviors manifest in groups. The ultimate goal is to develop robots that behave with the complexity of living beings so they can interact more realistically with actual animals.
The secret is imprinting. Around 5 hours after they hatch, chicks begin to grow deeply attached to their mother. It’s such a strong instinct that if something, anything, moves, chances are a chick will form a bond with it. That’s why farmers—at least the small-scale ones—go out of their way to bond themselves to their birds. It makes the critters more managable.
And researchers can use imprinting to trick chicks into falling in love with robots. First they put the chicks in little plexiglass boxes from which they watch PoulBot scoot back and forth. All the while the robot calls out, though not with pre-recorded hen sounds. “If you start to emit real sound, you have to understand what those real sounds mean, and you have to translate chicken language,” says Université Paris Diderot physicist José Halloy, co-author of a new paper detailing the process. So the robot makes sounds that are chicken-ish, which helps the creatures bond to it.
Now the chicks are ready to meet their adopted mother face-to-face in a little pen. PoulBot isn’t programmed to act like a classical chicken mom, though. Instead, it leads the chicks to a particular spot in the pen, constantly monitoring who’s following. “If someone is missing you have to go back and fetch them, stimulate the chicks to follow, and then go back to the target,” says Halloy.
An overhead camera tracks each chick, and PoulBot has a special covering around its base so the animals don’t get their toes squished in the tracks. (Tracks instead of wheels, by the way, so the works don’t get gunked up with chick crap. It’s a tank on a battlefield of excrement.) The researchers also programmed PoulBot with a behavior called avoid-running-over-chick. “If a chick has fallen asleep during the experiment and hence lies below the level of the sensors,” they write in their paper, they don’t want it to be in danger. PoulBot must not kill its fuzzy babies! So it uses accelerometer readings to tell if it’s no longer on flat ground, and will back up accordingly. “The results are not very interesting if you destroy half of your animals during your experiments,” says computer engineer and study co-author Alexey Gribovskiy of the École Polytechnique Fédérale de Lausanne in Switzerland.
Now, while the majority of chicks imprint on the robot, they imprint on it to different degrees, which is important because that influences the dynamics of the group. “Obviously if you have only strongly imprinted chicks, you get the military march,” says Halloy. “Everybody follows the leader. If you have a bunch of mixed weakly imprinted and strongly imprinted and in-between chicks, you have some kind of organized chaos there.”
Some chicks follow the robot and some chicks follow other chicks, creating a dynamic mob that’s tracked by the overhead camera. Algorithms even calculate their speed and acceleration, classifying every chick by how it’s behaving. This tells the researchers not only how well the robot is indoctrinating the subjects, but how chicks can vary in their acceptance of a fake mother.
Now, developing animal behavior models to power robots is hard. I can’t do it, and you probably can’t do it. “It takes a PhD to build a model, which means four years of work,” says Halloy. The PoulBot speeds that process up. “The idea was to use robots and artificial intelligence to automate as much as possible to produce a model faster,” Halloy adds. That’s right—postdocs aren’t safe from automation either.
Unravel the intricacies of flocking behavior and figure out what cues a robot needs to send to get an animal to accept it as a mother, and you can build robots that get animals to do certain tasks. “I could imagine scenarios where robots act to lead animals to a food source or a medical treatment area without stressing them,” says ecologist and biomimetic roboticist David Bierbach, who wasn’t involved in the research.
The shepherds on the farms of the future, then, may well be robots. Robots on tracks, not wheels, of course.