AutoCattloger: Automated Cataloging & Training

• AutoCattloger is an autonomous system that integrates sensing, conditional recording, and closed-loop behavior training for animals.
• It utilizes modular hardware (webcams, sensors, Arduino) and a Python control stack to synchronize stimulus delivery, response capture, and reward actuation.
• The system records only when specific triggers occur, reducing data redundancy and enhancing experimental efficiency.

Searching arXiv for the provided CATOS paper and adjacent AutoCattloger-related work to ground the article in cited sources. AutoCattloger denotes an autonomous cataloging, logging, and training paradigm in which sensing, event-triggered recording, stimulus delivery, response capture, reward actuation, and review tooling are integrated into a single system. In the CATOS line of work, the term is used for an automated cat behavior logging and training system built on the “CATOS: Computer Aided Training/Observing System” platform, whose stated purpose is continuous observation of irregular events, long-term intensive training, and training and testing of animals without human interference in order to eliminate potential cues and biases induced by humans (Oh, 2014). In later literature, the same spelling is reused for a dairy-cow catalog-construction front end, indicating that the term has also functioned as a label for broader automated cataloging and identification architectures (Lyu et al., 21 Aug 2025).

1. Historical setting and operational rationale

The CATOS implementation arose from three recurring needs in animal behavioral biology: observation of certain species continuously, or for documenting specific events which happen irregularly; longterm intensive training of animals in preparation for behavioral experiments; and training and testing of animals without human interference, to eliminate potential cues and biases induced by humans. The system was therefore designed to monitor visual and auditory streams continuously, record only when specified conditions are met, and automate the entire stimulus-response-reward loop so that subjects can engage the device at will without transport or separation (Oh, 2014).

This design tied observation and operant training together. CATOS did not merely record behavior; it also presented stimuli, accepted responses, and delivered rewards automatically. That coupling is central to the AutoCattloger concept in its original feline sense: logging is not passive archiving, but a condition-triggered, timestamped representation of behavior embedded within an experimental control system. A plausible implication is that AutoCattloger is best understood as a closed-loop behavioral instrumentation stack rather than as a simple video logger.

2. Core architecture of CATOS

CATOS combined commodity sensing devices, Arduino-mediated actuation, and a Python control program named AA (“Agent for Animals”). The hardware and software were explicitly modular and open-source, so that the same platform could serve both as an observer and as an operant trainer (Oh, 2014).

Layer	Components	Function
Sensing	Two webcams, microphone, temperature sensor, photocell, piezoelectric sensor	Visual, auditory, environmental, and reward-confirmation input
Actuation and animal I/O	Speakers, light bulb, fans, automatic feeder, three pushbuttons or touchscreen	Stimulus playback, environmental control, reward delivery, subject response capture
Control stack	AA Python program, OpenCV, NumPy/SciPy, wxPython, Matplotlib, Arduino serial module	Scheduling, event detection, multiprocessing, logging, visualization

The sensing arrangement included two webcams used simultaneously to monitor the experimental area for 8–12 hours/day for about 5 months, a microphone for auditory input including sound recognition triggers, and environmental sensors placed both in the enclosure and under the feeder. Actuation covered stimulus playback through speakers, automatic light control when ambient light dropped below a threshold, automatic fan control when internal temperature was high, and an Arduino-based feeder. Two feeder variants were described: an initial servomotor dispenser with an IR distance sensor for food estimation, later discarded due to inaccuracy, and a second feeder using a servomotor-driven Archimedes screw with a piezo sensor to confirm food emission events. The response interface likewise evolved from three physical pushbuttons to a touchscreen in the final product with improved functions (Oh, 2014).

AA used a multiprocessing architecture to avoid bottlenecks. Separate processes handled video-in per camera, video-out, audio-in, audio-out, schema logic, and inter-process communication through a message-board process. The Arduino module inside AA handled serial communication with the microcontroller. This separation of streams and control logic is significant because AutoCattloger depended on concurrency: monitoring, trial scheduling, recording, and reward control had to remain synchronized without blocking.

3. Event-driven logging and autonomous trial control

CATOS implemented continuous monitoring with conditional recording. Video and audio were recorded only when motion or sound triggers were detected, with thresholds configurable in the schema. During sessions, the output folder contained captured video frames and MP4 clips around events or trials, WAV audio files, movement-record CSV files summarizing blob tracking, a trial-result CSV file with correct/incorrect counts and per-trial records with timestamps, and a system log file containing process and device events. Post-session processing automatically generated movies, movement visualization JPEGs in which blob locations over time were encoded by circle color from black at the start to white at the end, and labeled sound files, all organized into timestamped subfolders in an archive directory (Oh, 2014).

The same software stack controlled the training protocol. In the proof-of-concept task, three distinct human speech sounds were presented, each associated with one of three pushbuttons in Build 1, or with corresponding touchscreen targets in Build 2. Trials were self-initiated: cats approached the device and trials could be started without human intervention, with a playback sound serving as a possible trial-start cue. The schema process selected a stimulus, enabled the response window, evaluated whether the pressed button matched the assigned sound, and delivered immediate positive reinforcement on correct trials by rotating the Archimedes screw and checking the piezo sensor for reward confirmation. Incorrect responses produced no reward. Inter-trial intervals were not explicitly specified in the paper and were typically controlled by the schema process; this suggests that ITI duration was parameterizable rather than fixed (Oh, 2014).

Review was also formalized. AA DataViewer loaded logs, CSVs, MP4, WAV, and movement JPEGs per session, presented quick visual summaries for rapid session assessment, and produced across-session performance graphs by selecting multiple archive subfolders. The trainer therefore reviewed behavior and training progress without attending sessions directly. A common misconception is that autonomy implies zero human involvement; CATOS instead removed experimenter presence during trials while retaining human maintenance, parameter tuning, and periodic review.

4. Pilot feline experiment and quantitative interpretation

The CATOS system was built and tested in a pilot experiment in which cats were trained to press three buttons differently in response to three different sounds, specifically human speech sounds, to receive food rewards. The overall system was in use for about 6 months, and training and observation occurred for about 5 months from mid-Oct 2012 to mid-Mar 2013. Two domesticated cats were trained. Both learned to approach the feeder upon playback sounds to obtain food, and one cat further learned the button-press contingency: pressing one specific button out of three to obtain a reward with over 70 percent correctness. The technical summary also states that the most recent performance graph shows overall accuracy greater than 70% on the three-sound/three-button task, and that each button’s accuracy exceeds the 33.3% chance level for a three-choice task, while ongoing training for full three-sound/three-button mapping continued (Oh, 2014).

The paper framed performance with standard categorical-task measures. Overall or per-class accuracy was defined as

$A = \frac{N_{\text{correct}}}{N_{\text{total}}}.$

For a three-choice task, the chance level was

$p_0 = \frac{1}{3} \approx 0.333.$

To test whether observed performance exceeded chance, the technical summary gave the one-sided binomial tail probability

$$P = \sum_{i=k}^{N} \left[ C(N, i) \cdot (p_0^i) \cdot (1 - p_0)^{(N - i)} \right],$$

with $k = N_{\text{correct}}$. It also reported an approximate normal confidence interval for accuracy,

$\hat{p} \pm z \sqrt{\frac{\hat{p}(1-\hat{p})}{N}},$

and noted that a Wilson score interval is recommended for better coverage at small $N$ or extreme $\hat{p}$ (Oh, 2014).

CATOS also quantified data-efficiency gains from event-driven recording. For a 15-day sample, total observation time was 406,138 s, or 112.8 hours. Recorded frames numbered 206,024; at an average 7.5 FPS, this yielded recorded video duration of about 27,470 s, or 7.6 hours. The recording fraction was therefore about 6.7% of the observation period. Over about 5 months, movement records, MP4, JPEG movement visualizations, and WAV files totaled 37.35 GB. In this context, AutoCattloger was not only a trainer but a filter on raw observability, drastically reducing irrelevant data while retaining irregular events (Oh, 2014).

5. Reliability, limitations, and replication constraints

CATOS incorporated several robustness measures. Reward confirmation via a piezoelectric sensor reduced silent reward failures: if the servo actuated but no vibration was detected, the system could flag the event or retry dispensing. Conditional recording minimized storage and focused attention on relevant events. Environmental automation through photocell- and temperature-based thresholds standardized lighting and cooling without manual intervention (Oh, 2014).

The limitations were equally explicit. Feeder failures could occur when motor activation was shorter than 0.5 s, and dry food geometry could interlock and block emission. This motivated the second feeder design based on an Archimedes screw and piezo feedback. The IR distance sensor used in the first build proved inaccurate for food-level estimation and was removed. Motion and audio thresholds required tuning to avoid false triggers from noise or non-relevant movements. The hardware and software were described as in alpha, with further development required for routine cognitive experimentation. Species generalization depended on appropriate input/output devices, enclosure design, reward type, and safety considerations, even though cost and complexity were kept relatively low through open-source software and commodity components (Oh, 2014).

Replication was facilitated by explicit build guidance and by software release under GPLv3 at https://github.com/jinook0707/CATOS_alpha. The specified assembly path included a computer with USB ports, two webcams, microphone, speakers, an Arduino microcontroller, motor shield, servomotor, Archimedes screw feeder, piezoelectric sensor, temperature sensor, photocell, fans, light bulb, protective housing, and either three pushbuttons or a touchscreen. The software environment required Python with OpenCV, NumPy/SciPy, wxPython, and Matplotlib, followed by configuration of the stimulus set, motion and audio thresholds, feeder actuation times, piezo sensitivity, environmental thresholds, session duration, and archiving options, then dry runs without animals to verify playback, input detection, dispensing, logging, archiving, and failure handling (Oh, 2014).

6. Extension of the idea beyond cats

Later work broadened both the term and the architecture. In dairy computing, “AutoCattloger” was used as the front end of a three-part system that builds a Cattlog of dairy cows in a herd with a single input video clip per cow. In that system, Mask R-CNN and HRNet localized masks and keypoints, aligned the cow to a canonical template, and converted top-view coat patterns into a compact 2048-bit Cow Barcode. Catalog entries were formed by taking the bit-wise statistical mode across all valid frames in the input clip, and identity was assigned by Hamming distance during later retrieval from continuous streams. The reported evaluation found 31 of 36 isolated cows with correct predicted IDs, or 86%, and 84 correct retrievals with 47 misses in non-isolated walking, corresponding to a retrieval rate of 64% (Lyu et al., 21 Aug 2025).

Related systems in adjacent domains retained the same core logic: automated localization, structured extraction, provenance, and constrained matching. AICat extended DCAT v3 for machine-readable registration of AI systems under the EU AI Act, with explicit RDF classes and properties for systems, models, datasets, provider and deployer roles, policies, licenses, and release dates (Golpayegani et al., 2024). LLM-based pipelines have been evaluated for automatic extraction of DCAT-compatible metadata such as title, description, creator, language, spatial coverage, issued date, keywords, and theme labels from text-based sources, using zero-shot, few-shot, and fine-tuned settings (Busch et al., 4 Jul 2025). In archaeology, AutArch automated object detection and recording from legacy PDF find catalogues through Faster R-CNN, contour extraction, OCR, and interactive validation, while a later vision-language pipeline formalized document localization, block comprehension, and foreign-key plus distance-based bipartite matching for archaeological catalog collection (Klein et al., 2023, Pang et al., 2024).

Astronomical work generalized the same pattern at survey scale. One line separated detection from catalog production, with NoiseChisel producing labeled “Objects” and “Clumps” images and MakeCatalog consuming those labels plus Sky and Sky-$\sigma$ images to compute fluxes, magnitudes, shapes, and flags (Akhlaghi, 2016). Another line made catalogs lineage-centric rather than permanently materialized, so that requested source sets and attributes drove automatic discovery, creation, optimization, and partial processing of astronomical Source Collections (Buddelmeijer et al., 2011). The DES Science Portal operationalized a related idea for “science-ready catalogs,” recording each step of catalog creation in a relational database and automating region selection, object selection, value-added joins, and standardized characterization for large survey analyses (Neto et al., 2017).

This suggests that AutoCattloger is not a single fixed software artifact. In the narrow sense it refers to the CATOS-derived automated cat behavior logging and training system; in the broader sense emerging from later literature, it denotes a family of autonomous cataloging systems that combine selective acquisition, structured representation, and machine-assisted or fully automated downstream organization.