The forest touchscreen that lets wild monkeys start the experiment

Trompudo took three days to approach the touchscreen. In Carol Clark’s account for Emory University, the wild capuchin first struck the back of the box, then touched the screen and received a piece of banana. A contraption assembled by researchers had become something the monkey could operate for himself.[1]

CapuchinAI, developed by an Emory and Georgia Tech team, was tested at Costa Rica’s Taboga Forest Reserve. Its camera, touchscreen and reward mechanism bring controlled learning tasks into the monkeys’ home environment.[1]

That arrangement gives the AI a role inside the experiment. The public software combines live detection and a reward interface; the animal supplies the action. A programmable screen lets researchers change the task without constructing a new physical puzzle each time.[4]

A camera connected to a banana dispenser

Clark describes a portable wooden station with a Raspberry Pi computer, webcam, screen and food dispenser. Its batteries could support eight hours of operation. Those mundane choices matter in a field instrument: computing has to fit around carrying the apparatus, powering it and keeping the reward mechanism working.[1]

Clark’s report describes the initial lesson: a blue screen invites a touch, and a Python script signals the motor to dispense food. The recognition software controls activation when a capuchin approaches. The animal learns to operate that connection.[1]

Individual recognition and field operation were separate achievements. A model trained to distinguish six monkeys performed strongly on held-out images. The deployed pilot instead recognised capuchin faces as a category, allowing monkeys outside that original six to participate. The published paper identifies autonomous tasks tailored to each individual as ongoing development.[2][3]

Who learns—and who gets a turn

During the two-week presentation to two habituated groups, 16 monkeys interacted voluntarily. Ten triggered rewards, and eight developed a robust association between touching the screen and receiving food. This established that the animals could learn the basic interface; it was a feasibility trial for cognitive testing.[2]

The preprint treated anticipation of the dispenser, before food arrived, as evidence of a learned association. A touch log alone cannot make that interpretation.[3]

Participation also reflected the monkeys’ social world. The preprint reports that dominant adult and subadult males accounted for much of the interaction, while some females were displaced. A voluntary test therefore samples opportunity and willingness to approach alongside whatever ability the task measures.[3]

The public repository separates model training from the Raspberry Pi software that runs detection and the reward interface. This makes the experimental arrangement inspectable: another team can see how an image recognition decision becomes a recorded visit or a dispensed reward.[4]

The appealing scientific question is now unusually concrete. Once a monkey understands how the station works, what can a carefully chosen screen task reveal about its learning in the social setting where it lives? The first useful result was persuading the animals to operate the instrument at all.

Report an error or suggest a correction