An ethogram is a comprehensive catalog of all the distinct behaviors a species performs, with each behavior clearly defined so that different observers can identify it consistently. Think of it as a dictionary for animal actions: rather than vaguely noting that a horse “looks uncomfortable” or a mouse “seems anxious,” researchers build an ethogram that spells out exactly what each behavior looks like in physical terms, such as “ears pinned flat against the skull” or “repeated circling along the cage wall.” The tool is foundational in animal behavior science, used everywhere from zoos and dairy farms to neuroscience labs and wildlife conservation projects, and its influence has grown as machine learning starts automating what human observers once had to do by hand.
What Goes Into an Ethogram
At its simplest, an ethogram is a list of behaviors paired with definitions precise enough that two strangers could watch the same animal and agree on what it is doing. Each entry names a behavior and describes the observable body movements involved, without interpreting what the animal “intends.” A recent analysis of livestock ethograms found consistent patterns in how behaviors are defined: feeding and drinking entries tend to be short, around six to ten words, specifying a body part, an action, and the object of that action (e.g., “mouth contacts feed in trough”). Resting and moving behaviors run slightly longer, roughly six to fifteen words, with resting defined by actions and objects and movement defined almost entirely by action words alone, like “walks forward with steady gait.”1PubMed Central. Foundations of Livestock Behavioral Recognition: Ethogram Analysis of Behavioral Definitions and Its Practices in Multimodal Large Language Models
Building an ethogram for a species that has never had one is a labor-intensive, iterative process. A team studying odor-exploration behavior in dairy heifers, for example, started with an initial list drawn from published literature and then expanded it through repeated rounds of video coding across forty separate tests. Behaviors were sorted into odor-directed exploration (sniffing, licking, biting) and non-odor exploration, each given tight operational definitions. The resulting ethogram included not just obvious actions like sniffing and scratching but also subtler signals like ear postures, tail postures, tongue movements, and head shaking.2PubMed. Development of an ethogram to investigate cattle odor exploration behavior of Swedish dairy heifers A researcher studying a different species or a different question would end up with a very different list, but the underlying discipline is the same: watch, define, test, refine.
One distinctive ethogram example comes from cephalopod research. A team cataloging the Mexican four-eyed octopus (Octopus maya) under laboratory conditions identified twenty-three distinct behaviors sorted into six categories, including feeding, locomotion, resting, den maintenance, and protection. Because so little prior literature existed on this species, the descriptions were almost entirely structural, focusing on what the body physically does rather than what function the behavior serves. Their data revealed that the octopus spends the vast majority of both daytime and nighttime resting, but daytime activity levels were significantly higher, with the most frequently observed active behaviors being grooming, crawling, and climbing.3bioRxiv. Specific Ethogram of the Mexican four-eyed octopus: Octopus maya
Why Definitions Need to Be Airtight
An ethogram is only as useful as the agreement between the people using it. If one observer scores “resting” every time a cat sits still and another scores “vigilance” because the ears are upright, the resulting data are noise. This is why ethogram development nearly always includes a formal reliability step, where multiple observers independently score the same footage and their agreement is measured statistically.
In a study of captive red pandas in zoos, four independent raters scored video footage using a draft ethogram, then met to discuss every disagreement and tighten the definitions. That cycle of scoring, discussion, and revision was repeated until the raters consistently agreed at a high level. The researchers set a threshold for agreement and did not begin final data collection until they cleared it, ensuring that the variation in their dataset reflected actual differences in animal activity rather than differences between observers.4PubMed Central. Applicability of Machine Learning in Behavioural Monitoring of the Red Panda (Ailurus fulgens) in Zoos The dairy heifer ethogram mentioned earlier went through the same kind of validation, achieving moderate to excellent agreement across observers, which confirmed that the definitions were clear enough to apply consistently.2PubMed. Development of an ethogram to investigate cattle odor exploration behavior of Swedish dairy heifers
This reliability testing is what separates a rigorous ethogram from a casual list of things an animal does. Without it, different labs studying the same species can end up with incompatible datasets, making it difficult to compare results or replicate findings.
Standardization Across Studies
One persistent frustration in the field is that ethograms are often built from scratch for every new project, even within a single species group. A researcher studying aggression in lions may define behaviors differently than a researcher studying play in snow leopards, and both may use terminology that does not map onto what a third team studying cheetah locomotion has published. This fragmentation makes it hard to pool data across studies or run meta-analyses.
To address this, some researchers have proposed standardized ethograms for entire taxonomic families. A standardized ethogram developed for the Felidae (the cat family) introduced the concept of “base behaviors” that remain consistent across all species, paired with predefined modifiers that individual studies can attach to suit their specific questions. The idea is that everyone studying any cat species uses the same foundational vocabulary, preserving comparability while still allowing flexibility for particular research goals.5Applied Animal Behaviour Science. A standardized ethogram for the felidae: A tool for behavioral researchers These standardized frameworks have not been adopted universally, but they represent a clear direction the field is heading.
Ethograms in Animal Welfare Assessment
One of the most practical applications of ethograms is in evaluating whether captive animals are thriving or suffering. Because ethograms quantify behavior in objective terms, they provide a way to track changes over time: a zoo animal that gradually shifts from a diverse behavioral repertoire to repetitive pacing can be flagged before the problem becomes severe.
Research across a wide range of species, from giant pandas and elephants to parakeets and chimpanzees, has shown that when stereotypic behaviors (repetitive, seemingly purposeless actions like pacing or swaying) increase, behavioral diversity drops. The pattern holds in both directions: pharmacologically induced stereotypic behavior also reduces the range of other behaviors an animal performs.6PubMed Central. Behavioral Diversity as a Potential Indicator of Positive Animal Welfare Ethograms make this measurable. Rather than relying on a keeper’s intuition that an animal “seems off,” welfare teams can track exact proportions of time spent in different behavioral categories and detect shifts that correlate with poor welfare.
A study of captive black rhinoceroses in a UK zoo compared traditional ethogram data with a different assessment method called Qualitative Behavioral Assessment, where observers rate animals on descriptive scales like “lively,” “nervous,” or “angry.” The researchers found meaningful overlap: agonistic behaviors such as horn clashing aligned with descriptors like Angry and Startled, while playful actions like head flinging matched Lively and Excited. Importantly, certain ethogram behaviors correlated in informative ways with the qualitative impressions: naso-nasal greeting and environmental investigation correlated with Active and Interested, while tactile contact negatively correlated with Angry and Nervous.7PubMed. Evaluating Qualitative Behavioral Assessment and Ethogram Techniques for Captive Black Rhinoceros (Diceros bicornis) This kind of cross-validation helps welfare scientists understand what their ethogram-based numbers actually mean in terms of an animal’s emotional state.
Diagnosing Pain in Ridden Horses
Ethograms have moved beyond the research lab and into clinical veterinary practice. The Ridden Horse Pain Ethogram is a striking example. It lists twenty-four specific behaviors, the majority of which are at least ten times more likely to be observed in lame horses than in non-lame horses. When a horse displays eight or more of these behaviors during a ridden assessment, it is likely experiencing musculoskeletal pain, even if the rider, trainer, or veterinarian cannot see obvious lameness. A key piece of evidence supporting the tool is that when lameness is relieved using diagnostic nerve blocks, the ethogram scores drop markedly, establishing a direct causal link between pain and the behaviors listed.8Equine Veterinary Education. The Ridden Horse Pain Ethogram
This kind of ethogram is designed to be usable by people outside the research world. Riders, trainers, and equine professionals can learn the twenty-four behaviors and apply the scoring system themselves, catching pain problems that might otherwise go unnoticed for months or years. It is one of the clearest examples of ethogram-based science producing a tool with direct, everyday utility.
Machine Learning and Automated Behavior Classification
Watching hours of video and manually scoring every behavior is slow, expensive, and mentally exhausting. It is also the bottleneck that limits the scale of behavioral studies. Over the past decade, machine learning has started to change that by automating the process of turning raw footage into ethogram data.
DeepEthogram, a software pipeline designed to be usable across species and recording setups, uses neural networks to analyze raw video pixels and classify behaviors frame by frame. In tests on mice and flies, it achieved above 90% accuracy on individual frames, matching the performance of trained human observers.9PubMed Central. DeepEthogram, a machine learning pipeline for supervised behavior classification from raw pixels That last detail is worth pausing on: human experts do not agree with each other 100% of the time either, so matching human performance is effectively the ceiling for any automated system working from the same data.
A related approach involves first estimating an animal’s body posture, then using those pose data to classify behaviors. SuperAnimal is a set of pretrained pose estimation models that work across more than forty-five species without requiring additional manual labeling. When researchers do need to fine-tune the models for a particular species, the process requires ten to one hundred times less training data than earlier approaches.10Nature Communications. SuperAnimal pretrained pose estimation models for behavioral analysis Separately, researchers have combined the pose estimation tool DeepLabCut with a behavioral classifier called SimBA to automatically score complex behaviors in rats during drug self-administration experiments, demonstrating that machine learning can handle operant conditioning tasks that were previously scored by hand or by simple lever-press counts.11PubMed Central. Analysis of Operant Self-administration Behaviors with Supervised Machine Learning: Protocol for Video Acquisition and Pose Estimation Analysis Using DeepLabCut and Simple Behavioral Analysis
One emerging frontier involves AI systems that do not rely on a predefined ethogram at all. A recent study used self-supervised learning to identify clusters of mouse behavior from three-dimensional body-posture data, then showed those clusters could detect depressive-like behavioral patterns. The system’s behavioral clusters were slightly shorter in duration than manually defined behavior categories but captured nuances that human-designed ethograms may miss.12PubMed Central. AI-driven decoding of naturalistic behaviors enables tailored detection of depressive-like behavior in mice This approach raises an interesting philosophical question: if a machine discovers meaningful behavioral units that no human ethologist ever defined, does the resulting output still count as an ethogram? The field has not settled on an answer, but the practical value of discovering new behavioral structure in large datasets is clear.
Accelerometers and Bio-Logging
Not all animals can be watched on camera. For species that roam vast distances, live underground, or are active only at night, researchers increasingly rely on small body-mounted sensors, especially accelerometers, to record movement patterns that can be mapped to ethogram categories after the fact.
The process requires its own validation step. A team studying African ground pangolins, a notoriously difficult species to observe because of their nocturnal, solitary habits, matched accelerometer readouts to behaviors confirmed by simultaneous video observation. Using a random forest classifier, they could distinguish five discrete behaviors (walking, digging, feeding, investigating ground, and stationary) with 85% accuracy, and three broader activity levels (low, medium, and high) with 94% accuracy. They then deployed the validated system on three free-ranging pangolins for several days to collect behavioral data in the wild.13Animal Biotelemetry. Classification of African ground pangolin behaviour based on accelerometer readouts: validation of bio-logging methods
A broader benchmarking study of machine learning methods for bio-logger data explored whether deep neural networks outperform older approaches like random forests when working with accelerometer data, and whether models trained on one species can be transferred to another. The researchers found that self-supervised pre-training on large unlabeled datasets can reduce the amount of hand-labeled training data needed, which is a significant practical benefit given how expensive annotation is for wildlife studies.14PubMed Central. A benchmark for computational analysis of animal behavior, using animal-borne tags For species where capturing and tagging animals is stressful and opportunities are rare, needing less labeled data per deployment is a real win.
Crowd-Sourcing Behavioral Observation
A less obvious approach to scaling ethogram-based research is recruiting large numbers of untrained observers through online platforms. This sounds risky, given how much effort goes into ensuring agreement among trained raters, but some researchers have found it surprisingly workable.
A video rating tool tested across three waves of crowd-sourced observers achieved strong agreement, with reliability scores climbing as high as 0.99 after removing ratings where the video had failed to load. The number of raters needed to reach stable agreement varied by study, ranging from about ten to seventeen, depending on how variable the behaviors being scored were.15PubMed Central. Using a new video rating tool to crowd-source analysis of behavioural reaction to stimuli Crowd-sourcing works best for behaviors that can be rated on simple scales, like how much an animal reacts to a stimulus, rather than for fine-grained ethogram coding that requires distinguishing between closely related behaviors. It is a supplement to expert coding, not a replacement, but one that dramatically expands the volume of data a lab can process.
The Anthropomorphism Trap
One challenge that hangs over every ethogram is the temptation to describe animal behavior in terms of human emotions or intentions. Calling a behavior “playful” or “frustrated” sneaks in an assumption about what the animal is experiencing, which may or may not be accurate. An ethogram is supposed to be purely descriptive: the animal did X with body part Y in context Z. Researchers have debated the limits of anti-anthropomorphism for decades. Taken too far, the insistence on mechanical descriptions can blind scientists to genuine cognitive and emotional capacities in animals. Taken too loosely, it introduces bias that is hard to detect.16PubMed Central. Anti-anthropomorphism and Its Limits
In practice, most modern ethograms strike a middle ground. The behavioral definitions themselves stay physical and observable. Categories that group those behaviors, like “agonistic” or “affiliative,” carry a light functional interpretation, but one grounded in known outcomes rather than assumed feelings. As the rhino welfare study showed, it is possible to correlate ethogram data with qualitative emotional descriptors after the fact, letting the data bridge the gap rather than smuggling assumptions into the definitions from the start.
Aquatic Species and High-Throughput Phenotyping
Ethogram-based approaches are expanding rapidly into aquaculture and aquatic biology, where behavior has historically been harder to track. A platform designed for high-throughput behavioral phenotyping of shrimp achieved tracking accuracy well above that of existing deep learning methods, outperforming algorithms like Track-RCNN and DeepMOT on standard tracking benchmarks.17ScienceDirect. HTBPPS: A high-throughput behavioral phenotyping platform for shrimp For an industry that manages billions of animals in dense populations, automated behavioral monitoring has obvious commercial appeal: detecting disease outbreaks, measuring feeding efficiency, and identifying stressed individuals before productivity drops. Ethograms designed for aquatic invertebrates tend to be simpler than those for mammals, often focusing on locomotion speed, feeding rate, and position in the water column, but the foundational logic is the same. Define the behaviors clearly, ensure observers or algorithms can identify them reliably, and collect data at scale.