Bridging Controlled Labs and Unpredictable Habitats
Primate brains didn't evolve inside sterile, temperature-controlled laboratory cages. They evolved in complex, competitive rainforests where survival hinges on social negotiation, spatial memory, and split-second decisions. Yet for decades, cognitive scientists faced a frustrating compromise. If you wanted precise experimental control, you had to bring animals into a lab, stripping away their natural social groups and life context. If you observed them in the wild, you got rich ecological reality, but almost no way to run controlled, repeatable tasks on individual animals.
Marcela Benítez, an assistant professor of anthropology at Emory University, decided that compromise wasn't good enough. Working alongside computer scientists and evolutionary biologists, her team set out to bring lab-grade precision directly into the canopy. Their solution—an automated, battery-powered platform called CapuchinAI—merges computer vision, interactive touchscreens, and brain-inspired edge computing into a single field-deployable unit. Tested on wild white-faced capuchins (Cebus capucinus / Cebus imitator) in Costa Rica's Taboga Forest Reserve, the system runs cognitive experiments without human handling or constant field intervention.
Hardware Architecture and DIY Garage Construction
Deploying delicate electronics into a humid tropical forest filled with inquisitive wildlife requires serious practical engineering. You can't just set a consumer tablet on a tree stump and expect it to survive tropical rain or twenty-pound monkeys trying to pry it open.
Benítez and Emory PhD student Federico Sánchez Vargas designed and constructed the physical prototype in a home garage. They sourced pine boards, heavy-duty deck sealant, rubber insulating strips, and plastic piping from Home Depot to build a weather-resistant, 20-inch wooden housing. Inside this enclosure sits a Raspberry Pi microcomputer—about half the size of a standard smartphone—connected to a wide-angle webcam, a low-cost interactive touchscreen, and a custom 3D-printed plastic food dispenser developed with Emory's TechLab.
Power comes from a lightweight, rechargeable battery pack that runs the entire setup for eight hours straight. When a capuchin completes a touchscreen task correctly, a small rotary motor turns the dispenser to drop a slice of dried banana into a tray. The entire rig cost remarkably little to build, proving that cutting-edge field research doesn't require five-figure custom hardware budgets.
Vision Pipelines and Simplified Python Control
The secret to running automated experiments on free-roaming animals is knowing exactly who is interacting with the machine at any given moment. Without individual identification, you can't deliver personalized task progressions or track learning curves over time.
To tackle identity tracking, Jacob Abernethy, an associate professor of computer science at Georgia Tech, and former graduate researcher Sai Rakshith Potluri adapted the open-source YOLO (You Only Look Once) neural network architecture. Emory and Georgia Tech undergraduates spent hundreds of hours framing monkey faces with bounding boxes across high-resolution GoPro footage. Trained on six target white-faced capuchins, this vision pipeline achieved 97 percent facial recognition accuracy across static images and live field video.
Scaling that pipeline to a full troop of 100 wild monkeys presented a practical bottleneck: the team didn't have high-quality training video for every individual in the forest. To solve this, Sánchez Vargas wrote a clever pair of integrated Python scripts running concurrently on the Raspberry Pi.
The first script simplified the initial vision filter so it recognized capuchins as a general species rather than classifying specific identities immediately. Whenever any monkey approached the box, the script triggered the webcam to record high-resolution video clips. That automated data collection built up a fresh image library for unindexed monkeys. The second script used Pygame to render interactive stimuli on the touchscreen and signal the reward motor when touched.
Deploying in Taboga Reserve and Beating Monopolization
Field testing took place over two weeks in northeastern Costa Rica at the Capuchinos de Taboga research site, a long-term field project co-directed by Benítez since 2017 with the Universidad Nacional Técnica of Costa Rica. The daily routine started early. Researchers woke up at 4:00 AM, hauled the heavy equipment down unpaved forest tracks, and mounted the box onto a wooden field platform.
For the first two days, no monkeys approached the device. Field work requires patience, and unaccustomed hardware can draw suspicion. But on the third morning, a dominant male capuchin nicknamed Trompudo ("big snout") picked up the scent of dried banana. He climbed onto the box, investigated the back, and eventually slapped the touchscreen. A banana slice dropped from the dispenser. Trompudo ate the treat and tapped the screen again, instantly grasping the connection between screen contact and food delivery.
Within days, 16 capuchins regularly visited the apparatus. Ten triggered food rewards, and eight established robust screen-reward associations without any human coaching. The physical enclosure also survived aggressive tampering from non-target species, including coatis—cunning members of the raccoon family famous for raiding manmade containers.
Automated field testing creates another challenge: social dominance. In wild primate troops, high-ranking males often monopolize prime food sources and drive lower-ranking individuals away. To ensure equitable data collection across the entire group, the team programmed anti-monopolization rules into the software. The Raspberry Pi tracks how many rewards each recognized individual receives during a session. Once a capuchin reaches its reward quota, the system temporarily halts task presentation for that specific monkey. Frustrated high-ranking individuals move on, opening up the platform for lower-ranking or more hesitant group members.
Individual Variation and Open-Source Primate Research
The pilot deployment revealed clear cognitive and behavioral differences among individual monkeys. Some capuchins acted as rapid problem solvers, mastering touchscreen interactions almost immediately. Others explored the platform cautiously with their mouths, learning to kiss the display to trigger rewards. A third group acted as observant social learners, standing back to watch group mates operate the box before stepping up to try it themselves.
Published as a proof-of-concept in the American Journal of Primatology under the title CapuchinAI 1.0, the project provides complete hardware blueprints and software source code. The authors intentionally designed the system using off-the-shelf components and open-weight AI models and software so researchers worldwide can adapt the platform for other species, from baboons to wild crows.
Funding for the development came from Emory's AI.Humanities program, the National Institute on Drug Abuse at the U.S. National Institutes of Health, the National Science Foundation, the Lewis and Clark Fund for Exploration and Field Research, and the Emory Center for Mind, Brain and Culture.
By combining decades of rich observational data on wild troop dynamics with automated, real-time experimental tasks, CapuchinAI carries forward the scientific ethos of the late primatologist Frans de Waal. De Waal insisted that understanding animal intelligence requires recognizing primates as distinct individuals with unique life histories. With battery-powered vision boxes running in the canopy, cognitive science can finally test those individual minds in the environments that shaped them.
