CapuchinAI: Open-Source AI Platform for Automated Cognitive Testing of Wild Primates
Researchers at Emory University and the Georgia Institute of Technology have unveiled CapuchinAI, an open-source, battery-powered platform designed to automate the cognitive testing of wild primates. By merging machine vision with interactive touchscreen tasks, the system brings laboratory-grade experimental control directly into the wild, eliminating the need for constant human intervention.
Field-tested in Costa Rica’s Taboga Forest Reserve, the system uses YOLO (You Only Look Once) computer vision to identify individual wild white-faced capuchins with 97 percent accuracy. It then delivers personalized cognitive tasks and dispenses food rewards automatically, marking a massive leap forward in field primatology.
What is artificial intelligence (AI) in this context?
Before diving into the mechanics of CapuchinAI, it’s worth understanding what artificial intelligence (AI) actually means here. As IBM defines it, AI refers to the simulation of human intelligence processes by machines—especially computer systems. These processes include learning (acquiring information and rules for using it), reasoning (using rules to reach approximate or definite conclusions), and self-correction.
In the case of CapuchinAI, AI isn't a sentient robot or a chatbot. It's a highly specialized machine learning model focused on one thing: recognizing specific monkeys in real-time video feeds. By leveraging a variant of the YOLOv7 architecture, the AI processes live webcam footage, draws bounding boxes around capuchin faces, and matches them against a database of trained identities. This is a classic example of embodied AI applied to a physical task—the AI's "body" is the Raspberry Pi, webcam, and touchscreen interface, and its "environment" is a dense Costa Rican rainforest.
Bridging the Lab-Field Divide
Primate brains didn't evolve in a sterile lab cage; they evolved in highly complex, unpredictable social and ecological landscapes. Yet, for decades, primate cognition has been studied almost exclusively in controlled laboratory settings. Lab experiments offer tight control over variables, but they strip away the natural social dynamics and varied life histories that shape how animals actually think.
Conversely, observing animals in the wild provides valid ecological context, but designing and controlling experiments in an unpredictable jungle is notoriously difficult.
Marcela Benítez, Emory’s assistant professor of anthropology and the project's senior author, wanted to bridge that gap. "The primate brain didn’t evolve in a lab," Benítez notes. "Yet primate cognition is rarely studied in the wild because the experimental control needed to measure cognition is difficult in unpredictable environments."
CapuchinAI solves this by bringing precise, lab-style experimental control directly into the wild. It records how natural environments shape individual cognitive strategies, allowing researchers to draw from decades of accumulated observational data on the life histories of individuals in the Taboga Forest Reserve.
The Hardware: Low-Cost, Wildlife-Proof, and Battery-Powered
The physical platform is a marvel of DIY ingenuity. Built primarily in Benítez’s garage, the system is housed in a weatherized, monkey-proof pine enclosure standing about 20 inches tall. The components are accessible and low-cost:
- Computing Core: A Raspberry Pi microcomputer (about half the size of an iPhone).
- Input/Output: A standard webcam and a low-cost touchscreen.
- Reward System: A 3D-printed plastic rotary food dispenser powered by a motor.
- Power: A lightweight battery pack that keeps the system running for approximately eight hours.
The entire box was sealed with deck sealant and rubber insulating strips to withstand tropical humidity and physical tampering. It even passed the "coati test"—coatis, members of the raccoon family notorious for breaking into man-made containers, were unable to breach the wooden enclosure.
Inside, the software runs two Python scripts simultaneously. The first handles the facial recognition model. The second uses a program called "pygame" to facilitate interactive stimuli for cognitive testing. When a capuchin approaches, the webcam triggers, and the system presents a simple stimulus—initially, a blue square on the touchscreen. Touching the screen signals the motor circuit to dispense a dried slice of forest banana.
Facial Recognition and the YOLO Model
The visual recognition backbone of CapuchinAI is its most technically ambitious component. The researchers, led by Georgia Tech’s Jacob Abernethy and former graduate assistant Sai Rakshith Potluri, trained an open-source model known as YOLO (You Only Look Once) on high-quality GoPro imagery of six wild capuchins.
Undergraduates at Emory and Georgia Tech performed the labor-intensive work of digitally placing "bounding boxes" around the faces of monkeys in thousands of still images and videos tagged with their identities. The result was a facial-recognition system capable of identifying these six capuchins with 97% accuracy from static images, video, and live footage in the field.
However, the system was designed to be scalable. As first author Federico Sánchez Vargas pointed out, "The model was great at identifying six monkeys, but there are 100 capuchins at the Costa Rica field site." To solve this, the system was programmed to record video whenever a capuchin approached, rapidly generating a larger, up-to-date dataset of faces. These videos are then used to continuously retrain and expand the facial-recognition model, allowing CapuchinAI to eventually identify the entire troop.
Crucially, Sánchez Vargas "dumbed down" the model's parameters so it recognized any capuchin monkey—even if it wasn't in the database—and ignored all other wildlife. This ensures the touchscreen only activates for the target species.
Automated Cognitive Testing in the Wild
The true breakthrough of CapuchinAI lies in its ability to deliver personalized cognitive tasks. Once the AI recognizes a specific monkey, the software presents a cognitive test tailored to that individual's "level" in the testing sequence. If the monkey is unfamiliar, the system defaults to the baseline habituation stimulus: touch the screen to get a reward.
The platform tests four broad domains of cognition:
- Learning speed: How quickly a monkey grasps the screen-reward association.
- Impulse control: Tests requiring the monkey to wait or inhibit a reflex.
- Cognitive flexibility: Tasks that require switching rules or strategies.
- Working and long-term memory: Evaluating short- and long-term retention of visual or spatial cues.
Preventing Monopolization and Ensuring Data Quality
In wild primate social structures, dominant individuals often monopolize resources. To prevent a single dominant capuchin from hogging the testing box, the automated software tracks individual participation limits per session. Once a monkey reaches its designated reward threshold, the touchscreen temporarily halts task presentation for that specific individual. This encourages lower-ranking or less assertive group members to interact with the device, ensuring balanced data collection across the troop.
Field Results: Rapid Habituation and Individual Variation
Upon arriving at the Capuchinos de Taboga research facility, the team faced a moment of anxiety. For the first two days, no capuchins visited. On the third day, a large male named Trompudo ("big snout") couldn't resist the scent of banana. He climbed the box, slapped the back of it, and eventually slapped the touchscreen. A banana slice popped out.
"Trompudo gobbled up the food then put his hand on the screen again. Another banana slice popped out," Sánchez Vargas recalled. "It was almost like you could see him realizing, 'Ah, that’s what you have to do, touch the screen!' It was amazing to watch an individual learn something so quickly."
Over a two-week pilot phase, 16 wild white-faced capuchins (Cebus imitator) voluntarily interacted with the apparatus. Ten triggered rewards, and eight formed and retained robust screen-reward associations. The researchers observed distinct learning styles: some were fast learners, others investigated the box with their lips (learning to "kiss" the screen for a reward), and some were late adopters who hung back to watch others figure out the apparatus before approaching it.
A Powerful New Tool for Primatology
"This project builds on the legacy of Frans de Waal," Sánchez Vargas noted. De Waal, who passed away in 2024, pioneered studies of animal cognition and treated primates as individuals. "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."
The research team emphasizes that CapuchinAI doesn't replace human researchers. "Frans de Waal said that you cannot study cognition if you don’t understand the animals," Sánchez Vargas adds. "Our AI methodology doesn't replace the need for human researchers in the field. It’s essential to have rich, observational datasets gathered by people working on the ground."
The open-source paper includes a complete guide to the computer coding, a blueprint for building the low-tech field platform, and instructions for integrating the components. The researchers hope other scientists will adapt this AI methodology to generate cognitive data spanning different species of wild primates in a range of environments.
Funding and Publication
The project received seed funding from Emory’s AI.Humanities program and was supported by the National Institute on Drug Abuse (R34DA061925), the National Science Foundation (BCS-2127373), the Lewis and Clark Fund for Exploration and Field Research, and the Emory Center for Mind, Brain and Culture.
The proof-of-concept, titled "CapuchinAI 1.0: Development of a Machine Learning-Based Touchscreen Paradigm to Test Cognition in Wild Capuchins," was published in the American Journal of Primatology (DOI: 10.1002/ajp.70194).
Source: AI Automates Cognitive Testing of Wild Monkeys - Neuroscience News (July 29, 2026)
