Wild monkeys got a touchscreen and 97% of the time the AI knew exactly who was tapping [Field Test]

It took three days and a large male capuchin named Trompudo, whose name means “big snout,” before anyone knew whether the box worked. He climbed on top of it, slapped the back panel, then slapped the touchscreen. A slice of dried forest banana dropped out.

He ate it. Then he put his hand on the screen again. Another slice.

“It was almost like you could see him realizing, ‘Ah, that’s what you have to do, touch the screen!'” said Federico Sánchez Vargas, an Emory Ph.D. student of anthropology and first author of the paper. “It was amazing to watch an individual learn something so quickly.”

That moment, in the Taboga Forest Reserve of northeastern Costa Rica, is the whole proof of concept behind CapuchinAI, a system published this month in the American Journal of Primatology by researchers at Emory University and Georgia Institute of Technology. The headline number: the facial-recognition model identified individual wild capuchins with 97% accuracy from static images, video and live footage in the field.

The lab problem nobody had solved

Primate cognition research has a control-versus-reality trade-off that’s been baked in for decades. Lab experiments are tightly controlled, but the results may be skewed because the animal isn’t interacting within its natural environment. Field experiments have valid social and ecological context and are miserable to design and control.

“The primate brain didn’t evolve in a lab, it evolved in complex, competitive environments,” said Marcela Benítez, Emory assistant professor of anthropology and senior author of the paper. “Yet primate cognition is rarely studied in the wild because the experimental control needed to measure cognition is difficult in unpredictable environments.”

“I’m trying to bridge that gap,” Benítez said.

The bridge, in this case, runs on a Raspberry Pi about half the size of an iPhone. It identifies an approaching monkey, presents a learning task tailored to that individual on a touchscreen and dispenses a food reward if the monkey gets it right. The whole rig runs for eight hours on a lightweight battery pack before it needs recharging.

The model was too good and that was the problem

Jacob Abernethy, Georgia Tech associate professor of computer science, and Sai Rakshith Potluri, a former Georgia Tech graduate research assistant now a software engineer at ExtraHop in Seattle, built the recognition model on YOLO, the open-source “You Only Look Once” software, and trained it to run on a laptop. Training data came from high-quality GoPro footage of six wild capuchins interacting with testing platforms at Taboga.

Undergraduates at both schools did the grinding part: drawing bounding boxes around monkey faces across thousands of images and videos tagged with individual identities. That labor produced the 97% figure.

It also produced a ceiling. Six monkeys is not a study population.

“The model was great at identifying six monkeys, but there are 100 capuchins at the Costa Rica field site,” said Sánchez Vargas, who joined Emory as a graduate student in 2023. “And we didn’t have high-quality video of all of these individuals, which is needed to train the model.”

His fix was to make the AI dumber

Sánchez Vargas holds undergraduate degrees in evolutionary biology and psychology. He’s not an expert coder. He went into the Python anyway and stripped the model’s parameters down.

“I essentially dumbed it down,” he said, so that instead of classifying individual capuchin faces, the system recognized any capuchin monkey and only capuchins.

The logic is a bootstrap. Any capuchin approaching the touchscreen triggers a webcam to start recording, which generates fresh, current video of monkeys the model has never seen. Feed that back into training and the recognition repertoire grows on its own.

A second Python script uses pygame, a program for building interactive stimuli in games and cognitive testing, to run the tasks. The habituation stimulus is deliberately crude: a blue square filling the touchscreen. Touch it, the script fires a motor circuit, banana comes out. Both scripts run simultaneously on the Pi.

Built in a garage, tested by raccoons

“We built it in my garage,” Benítez said.

She and Sánchez Vargas went to Home Depot for pine planks, deck sealant, rubber insulating strips and plastic piping. Emory TechLab handled a 3D-printed plastic food dispenser. The finished housing is a wooden box only about 20 inches (51 centimeters) tall, holding a webcam, touchscreen, Raspberry Pi, dispenser and rotary motor.

“A person helping us asked, ‘What is the project you’re working on?'” Benítez said. “We didn’t go into it; it would have been a bit complicated to explain.”

“It was a lot of work but also a lot of fun,” Sánchez Vargas said. “One of the best parts of working in animal cognition is getting creative, trying to put yourself into the mind of the animal you’re studying, so you can develop a good way to engage them in experiments.”

Durability testing wasn’t a spec sheet exercise. The box held up against capuchins trying to bust it open. The facial-recognition software keeps the system from activating for other wildlife, but that doesn’t stop coatis, raccoon relatives notorious for breaking into containers and houses, from trying to pry it apart. The box passed the coati test.

Getting there is the unglamorous part

“We had to get up super early, at four in the morning,” Sánchez Vargas said, “and drive a long way down this bumpy road to get to the research site.”

Then the waiting. “For the first couple of days that we did this, no capuchins visited,” he said. “I was beginning to worry.”

Trompudo broke the ice on day three. After that, more monkeys engaged. “Even I was surprised by their enthusiasm,” Sánchez Vargas said.

Sixteen monkeys, sixteen different learning curves

The pilot phase drew 16 capuchins, and the individual variation showed up immediately. Some are fast learners. Others take longer to work out the screen-then-food connection. A few investigated the box with their lips and learned to kiss the screen for a banana slice.

And then the late adopters, who hang back and watch their friends use the box before touching it themselves.

The team is retraining the facial-recognition model on video of those 16. Next come cognitive tests across four broad domains: learning, impulse control, cognitive flexibility, and short- and long-term memory. Recognize a known individual, serve the test matching that monkey’s level in the sequence. Unknown individual, serve the baseline blue square.

Machine vision lets the model swap tests between identified capuchins fast enough for multiple individuals to participate. There’s also a cap on how many rewards one animal can earn per session, which keeps a dominant capuchin from camping the platform. That’s a design detail that only comes from knowing capuchins.

What the data is actually for

Benítez co-directs Capuchinos de Taboga, a project launched in 2017 with the Universidad Nacional Técnica of Costa Rica. Her work sits across anthropology, psychology and evolutionary biology, covering both a captive population of tufted capuchins in a lab and the wild white-faced capuchins at Taboga. The site carries decades of observational data on individual life histories, which is what makes the automated testing worth doing.

“We can explore outstanding questions about how the environment, individual experiences and behaviors connect to cognitive abilities,” Benítez said. “Why are some individuals better at some tasks than others? How do different individuals adapt to different situations? How do different cognitive strategies relate to fitness?”

The paper is open source and includes the coding guide, a blueprint for building a low-tech, low-cost field platform, and instructions for wiring it all into a closed-loop pipeline. The authors want other researchers to adapt it for different wild primate species in different environments.

The part the AI can’t do

Sánchez Vargas frames the project as continuing the work of Frans de Waal, who directed Emory’s Living Links Center for the Advanced Study of Ape and Human Evolution, pioneered animal cognition research and wrote best-selling books that popularized the field. De Waal died in 2024.

Beyond the lab experiments, de Waal “gave us intimate, beautiful portraits of the lives of primates, treating them as individuals,” Sánchez Vargas said. “Our AI method allows us to more deeply understand individuals that we already have data on through field observation. We can now automate cognitive testing of them and quantify the findings. It’s a way of getting into the minds behind the personalities. Studying individuals in their natural environments, where there are tons of variations in their life experiences, lets us learn how environmental influences shaped them.”

He’s explicit about the limits. “Frans de Waal said that you cannot study cognition if you don’t understand the animals,” Sánchez Vargas said. “Our AI methodology doesn’t replace the need for human researchers in the field. It’s essential to have rich observational data sets gathered by people working on the ground.”

Which is the honest read on CapuchinAI. It’s a 20-inch pine box with a $35-class computer inside that automates the tedious half of a job whose valuable half still requires someone driving a bumpy road at four in the morning. “It’s a powerful new tool in the primatology toolkit,” Benítez said. The toolkit part matters more than the AI part.