Ethics statement

Approval for data collection protocols at Suaq was provided by the Indonesian State Ministry for Research and Technology (RISTEKDIKTI) and the National Research and Innovation Agency (BRIN). The reporting of this study meets ARRIVE guidelines.

ABM structure

We simulated the developmental trajectory of orangutan diet repertoires day by day for the duration of their immature period, up to the approximate age of first reproduction (15 years). Simulated immatures initially had no known food items in their diet repertoire. Each day, we generated the number of feeding patches a simulated immature would visit by sampling from a Poisson distribution, with a mean equal to the mean number of patches orangutan mothers at Suaq visit per day (27 feeding patches). For each feeding patch, we generated a single type of food item on the basis of the rates at which different foods were encountered at Suaq. Additionally, for each feeding patch, a simulated immature was assigned one of four social states: (1) alone, (2) distant association, (3) close association and (4) peering. These four states were chosen because wild immatures exhibit more frequent exploratory behaviours when closely associated with conspecifics44 and after peering38. In the wild, the probabilities of being in close association with a conspecific and engaging in peering, and the resultant likelihood of exploring food items in both of these contexts, all change over the course of development27,44,58. We therefore calibrated the probability that simulated immatures explored the food item in a feeding patch by both their age and their social state, using estimates from long-term data on wild orangutans. If a simulated immature performed an exploration behaviour, this exploration was logged in a list of explorations performed by the immature across their lifetime. At the point when a particular food item had been explored enough times to meet a minimum threshold for learning, this food item was added to the simulated immature’s diet repertoire (with the number of required explorations being scaled to the complexity of the food item’s associated processing behaviour). If a simulated immature encountered a food item that was already present in their diet repertoire (and therefore known), their behaviour was marked as ‘feeding’.

Our simulated immatures did not move around a simulated space (that is, our ABM is spatially implicit) but instead were presented with food items and social environments that reflect the opportunities presented to wild orangutans (akin to individuals being positioned in front of a ‘conveyor belt’ of food patches, which they are exposed to in turn, alongside a generated social state for each patch). This allowed us to calibrate our model precisely to reflect the opportunities afforded to immatures in the wild and circumvent unnecessary error introduced through estimations of movement rates across simulated environments. All steps to verify that our model operates following programmed protocols can be found in this Article’s associated code.

All ABM coefficients were calibrated using data from wild individuals. Further information about the statistical models we used to analyse data from wild individuals—including the rationale behind the choices of our models and their structure—can be found in Supplementary Information section 2 (alongside all model summaries and all resulting coefficients for the ABM; Supplementary Tables 2–5). The experimental treatments we applied to our ABM are outlined in Results, and for each experimental treatment, we ran 250 iterations (a compromise between large sample size and available computational power). When analysing data from wild individuals, we used all available data where possible (unless otherwise specified for a particular analysis), and these sample sizes are comparable to or larger than those of similar studies27,29,30,38,39,42,43,44,45,57,58,59,60,69.

Study site and long-term data collection

We calibrated our model using data on wild orangutan behaviours collected at Suaq Balimbing (South Aceh, Indonesia). Since 1994, data have been collected at Suaq through daily focal follows, where observers collect behavioural and association data using instantaneous scan sampling at 2-minute intervals. All-occurrence sampling of key behaviours is also performed during this time frame—for example, on peering behaviours (the full protocol for data collection can be found at https://www.ab.mpg.de/571325/standarddatacollectionrules_suaq_detailed_jan204.pdf). We sampled data collected between 2007 and 2019, including 2,676 follows on 132 individuals. When estimating parameters for our ABM, we used specific subsets of these data on the basis of the relevancy of the ages and social classes of focal individuals and, where necessary, the length of focal follows.

All individuals included in this study are well known and form part of long-term data collection at Suaq. The ages of immature individuals were estimated on the basis of known births or physical characteristics at their first encounter. Immature individuals are classed as ‘dependent immatures’ up until the age they are observed ranging independently from their mothers for at least two consecutive follows. After this age, immatures are classed as ‘independent’. Immatures reach independence at Suaq at between 7 and 9 years of age (mean, 8.1 years; N = 8; maximum age estimated, 9 years). We used the maximum age—9 years—as the threshold age for the onset of independence, thus offering a generous baseline for the latest reasonable age for adult-like diets to develop in the wild. Following the onset of independence, associations with mothers and other individuals become progressively less frequent35. Opportunities to benefit from other individuals’ diet knowledge thus become rarer, and immatures must increasingly rely on their own foraging knowledge for survival. Individuals are classified as adults once they reach the average age at which females first reproduce (15 years30; females who reproduced before this age are classed as adults from the age they give birth to their first offspring).

Whenever the focal individual was feeding or exploring a food item at a 2-minute scan, the food item was noted. Food items are recorded at Suaq as a combination of the consumed species and, where relevant, the part of the species being consumed. For plants, food items are differentiated according to the organ(s) being consumed (for example, ‘Leaf’, ‘Fruit’, ‘Pith’, ‘Bark’, ‘Flower’, ‘Seeds’ or the general vegetative material of the plant, ‘Veg’). For insects, we recorded the specific type of insect on the level of the family or clade (for example, all ants are coded as ‘Semut’) but did not differentiate on the parts of the body given they were often consumed whole. Food items were classified according to the complexity of the behaviour required to process the item before ingestion29 (on a ranked scale from 0 to 5; see below), including all steps to acquire edible material and dispose of waste. Items eaten whole (for example, many types of leaves) are marked at complexity level 0. Each additional processing step (for example, peeling or spitting out inedible material) increases the complexity score by one. Behaviours that involve the use of tools are marked at the highest complexity score, 5.

Adult diet-repertoire size

Because estimates of diet-repertoire size are highly dependent on observation time27, we estimated the total size of the diet repertoire of adults at Suaq by modelling the cumulative number of different food items consumed by specific adults as a function of sampling effort (measured as the number of behaviour scans). To estimate the total adult repertoire size, we fit a nonlinear mixed-effect model using the Michaelis–Menten equation70, with individual identity as a random intercept (the model was fit using the nlme package71 v.3.1.168). Using the Michaelis–Menten equation permitted us to estimate the asymptote in repertoire size as a function of sample effort (Vmax). This asymptote was taken to be representative of the total adult repertoire size.

Feeding patches

To estimate the mean number of feeding patches that wild dependent immatures visit in a day, we used the number of feeding patches visited by their mothers. Dependent immatures consistently follow mothers through their home range57,58 and are therefore exposed to a similar number of feeding patches—and array of different food items—per day. The number of feeding patches visited per day can be reliably estimated for mothers, as immatures may not feed if a food is not yet within their diet repertoire (that is, if it is unknown). We inferred the number of feeding patches that mothers visited via counting the continuous periods in which mothers consumed specific food items, as recorded in the focal follow data. We did not pay attention to breaks from feeding introduced by other behaviours that took place at feeding patches (for example, nursing or resting). However, if a mother began eating a different food item, this was classed as entering a new feeding patch. Our estimates include the possibility that mothers visit multiple feeding patches across a given day that contain the same food.

To model the average number of feeding patches visited by mothers, we constructed a Poisson GLMM to model the mean number of feeding patches encountered per follow, with focal ID as a random intercept.

To model the content of feeding patches, we estimated the probability of different food items being encountered across all follows for all adult orangutans. Including data from all adults ensured that our sample size was sufficiently large to yield reliable estimates of the probability of encountering all food items at Suaq, including foods that are rarely eaten. For each food item, we counted the number of follows (and therefore days) this food item was available for consumption (its food item frequency). By dividing each food item frequency by the sum total of all frequencies, we generated a discrete probability distribution for the probability of encountering different food items, which is proportional to how frequently each food is observed being consumed in the focal follow data (see the associated data for the full list of food items and their probabilities of being encountered). Each feeding patch contained a single type of food item; while different foods can occur in close physical proximity in the wild (for example, leaves and fruits from the same tree may be considered both in the same patch), restricting each patch to one food item closely resembled how we estimated the number of different feeding patches mothers encountered per day (where swapping between foods in close proximity would have been considered feeding in different patches).

Social states across development

If the simulated immature encountered a food item that was not in its diet repertoire, our ABM assigned that immature as being in one of four social states: alone, distant association, close association or peering. The social states of close association and peering relate to possible forms of social learning available to wild immatures (including processes of enhancement and observational social learning, respectively). The probability of a simulated immature being in each social state was determined by a hierarchy of decisions. First, is the simulated immature associated with a conspecific? If not, they are assigned the state alone for that specific feeding patch. Second, if associated with a conspecific, is the simulated immature in close association? If so, for that feeding patch, they are assigned the state close association; otherwise, they are assigned the state distant association. Third, if in close association, is the simulated immature peering? If so, they are assigned the state peering for that feeding patch.

We estimated the probabilities of simulated immatures being in each social state using focal follow data for wild individuals within their first 15 years of life. We also estimated their probability of exploring food items when in each social state across this developmental period (see ‘Exploration rates across development’).

Probability of association

To estimate the probability of a simulated immature being associated with a social partner, we counted the number of scans where a focal individual was within 50 m of at least one conspecific (and therefore in association) per follow, versus those in which they were outside this range and classified as alone. We calculated the probability of an immature being associated with a conspecific at different ages using a quasibinomial GLMM (with data points weighted by the total number of scans per follow and with focal ID included as a random intercept). At each age, the probability of being alone was 1 − pAge(associated).

Probability of distant or close association

For all scans where an individual was in association, we determined the probability of being in close or distant association. An analysis of our data showed that orangutan exploration rates when within 0–10 m of a conspecific were very similar across ages; however, at greater distances (10–50 m) exploration rates across ages were much lower. We therefore defined close association as being within 10 m of at least one conspecific. For each follow, we divided the number of scans where an individual was associated with a conspecific into those where they were in close and distant association. We then modelled the probability of being in close association when associated with a conspecific at different ages, using a quasibinomial GLMM (with data points weighted by the total number of scans for each follow and with focal ID as a random intercept). At each age, we set the probability of simulated immatures being in close association (if associated with another individual) to be equal to that for wild individuals, and pAge(distant association) was 1 − pAge(close association).

Probability of peering

When in close association, orangutans may perform close-range observation of the behaviour and/or objects manipulated by conspecifics (peering). To estimate how peering rates change with age, we used a quasibinomial GLMM comparing the number of close association scans where an individual was peering with the number of scans where an individual was not peering (weighted by the total number of scans in close association and with focal ID as a random intercept). We used this rate as the probability of simulated immatures peering when in close association in a feeding patch at a given age.

Exploration rates across development

The majority of great ape learning is facilitated through exploration, which can be mediated and enhanced through social interactions27,38,39,40. Exploration is defined as repetitive, often destructive manipulation (including failed feeding attempts) of objects (including food items), while the visual and tactile foci of the individual are on the object27,72. To quantify the effects of these different social factors on the probability of exploration, we modelled the probability of wild immatures exploring when in distant and close association, as well as after peering. It was not possible to accurately estimate the exploration rates of wild orangutans when alone at all ages, as dependent individuals spend virtually all of their time in association with their mothers. We therefore set simulated immatures’ exploration rates when alone to be equal to the exploration rate at distant association for each age.

Exploration in distant and close association

We estimated how the exploration rates of wild immatures change over development, including when in both distant and close association. For each follow, we sampled all behaviour scans where an immature was in association, and we counted the number of scans where the focal immature was performing an exploration behaviour versus scans describing any other behaviour. We partitioned these counts into those where an individual was in close and distant association. To control for the effect of peering (which can increase exploration rates; see below), we excluded exploration scans where an orangutan had peered at a food item of the same species within the previous hour. We controlled for peering at the level of the species of the food item, because if an immature is drawn to a species of food following peering, they may choose to explore multiple parts (for example, leaves or bark). We modelled the probability of exploration over age for both association categories (close and distant) using a binomial GAMM (with follow number and focal ID as random intercepts and with our model weighted by the total number of scans sampled in each association proximity condition per follow). We translated the model’s mean exploration probability at each age into the exploration rates of simulated immatures when in distant and close association.

Exploration following peering

To estimate exploration rates following peering, we used a subset of the follow data where observers recorded peering events on an all-occurrence basis. For each follow, we recorded each instance (or the last instance in case of peering events in close succession) when an immature peered at a specific species of food item, and whether the immature explored this food item within the following hour (coded as a binary Y/N). We determined whether peering had a significant effect on the exploration rate by comparing the probability of exploring a food item after peering (if the food item had not already been explored in the hour before) with the probability of exploring food items in the hour prior to peering events (a baseline for exploration in contexts that may elicit peering, such as around novel foods). For both of these scenarios, we used binomial GLMMs with focal ID as a random factor. We also estimated whether the effect of peering on exploration probability changed over the course of development. We modelled the proportion of peering events that were followed by exploration behaviours in the subsequent hour, compared with those that were not followed by exploration during the same time frame, for wild immatures across the first 15 years of development (binomial GLMM with focal ID as a random factor and weighted by the number of data points at each age). Model estimates were then translated into the probability of simulated immatures exploring in the peering state at each age.

Number of explorations needed to learn

The total amount of exposure wild immatures require to learn how to process a food item depends on the number of processing steps that must be performed before it can be ingested29 (herein considered a food item’s complexity). We therefore scaled the number of explorations that simulated immatures require to learn how to consume food items by their complexity.

To estimate the number of times wild immatures explored food items of each complexity category prior to learning how to eat them, we first sampled the five most common food items of each complexity category and determined the earliest age at which wild orangutans were observed consuming each food item using focal follow data (the earliest estimated learning age). We restricted this analysis to immatures who had been followed since at least the age of 1 year, ensuring that the first age that food items were observed being consumed was probably close to the age of learning. To estimate the number of times a species of food item was encountered before an individual learned how to process and eat it, we calculated the rate at which food items occurred in adult focal follows (their probability of being encountered each day). We multiplied this number by the number of days preceding the individual’s estimated age of learning (that is, the total number of days spent foraging prior to learning). This provided an estimate of the number of times a particular type of food item was encountered prior to learning. We then estimated the number of times an immature explored a food item prior to learning how successfully eat it, by multiplying the number of encounters by the mean exploration rate across the dependency phase.

Across the five most frequent food items of each complexity category, we averaged the estimated number of explorations required for learning. As more complex food items are rarely encountered, this analysis permitted us to estimate the required number of explorations for complexity categories 0–2 (foods of category 0 required one exploration, category 1 required two explorations and category 2 required two explorations). We extrapolated this trend to higher complexity categories by incremental increases of one exploration for every two increasing steps in complexity (three explorations for category 3, three explorations for category 4 and four explorations for category 5).

Reporting summary

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.