Insects / Machine-learning and follower-response study

Honeybee dance records connect movement with followers and food sites

Publication year 2023 · Note reviewed 2026-10-25

Dormagen et al. combined waggle-phase detection, individual tracking, follower observations and feeder records. The paper reports more than 100,000 detected waggle phases. A manually annotated subset covers 176 videos, 3,278,121 detections from 28,873 tracks and 4,389 bee IDs, including 1,107 waggle phases and 279 follower events. The Zenodo v1.0 API lists 11 deposited files totaling 4,870,358,178 bytes (4.87 decimal GB), alongside more than 46 million 2019 track detections. This is a rare data path connecting a well-defined sender signal to observed receiver behavior and known food-site outcomes.

What we reviewed · Targeted primary-paper and dataset-metadata review; no payload acquisition or reproduction

Targeted review of the PNAS Nexus abstract, communication background, colony and camera setup, annotation and hourly sample grouping, follower/feeder results, data statement and article license; on 2026-10-25, rechecked the complete Zenodo v1.0 API file manifest and dataset license. The 11 listed sizes sum to 4,870,358,178 bytes; the named follower CSV is 9,171,824 bytes. No data, video, code or model was downloaded, locally hashed, inspected or executed.

No independent Talk2Nature insect-behavior or ML/statistical specialist has reviewed this assessment.

Study at a glance

Scale: The Zenodo v1.0 API lists 11 files totaling 4,870,358,178 bytes (4.87 decimal GB); Berlin2019_followers.csv is 9,171,824 bytes. The paper reports more than 46 million 2019 track detections and more than 100,000 waggle phases.
Observed response: The paper links dance features to follower behavior and feeder experiments, giving the signal a measurable external outcome.
Independent units: The article’s sample split groups observations by hour; it does not establish transfer to held-out bees or a new colony.
Current decision: Promising communication-methods candidate, not a selected Talk2Nature benchmark. Published results, single-site scope, the follower-file intake limit, code rights and independent review remain gates.

Reading location: Dormagen et al. (2023), PNAS Nexus, DOI 10.1093/pnasnexus/pgad275: abstract, communication introduction, colony setup, sample-grouping method, Results, data availability and article license; Zenodo v1.0 record DOI 10.5281/zenodo.7928121 and official API file sizes/license rechecked 2026-10-25.

What this does — and doesn’t — establish

This is not a blank slate: the paper already analyzes waggle dance/follower relationships and feeder outcomes. It spans one observation-hive site across two seasons; the published sample-level cross-validation groups by hour rather than holding out bees or colonies. The 2021 recordings lack marked individual identities. The dataset is specialized tracking and video data, not an ordinary-phone audio corpus. The public record contains fine-scale feeder coordinates, which are not reproduced here. The separately named Berlin2019_followers.csv is 9,171,824 bytes, above the current 5 MiB per-file collector cap; smaller dance/feeder metadata files have not been shown to substitute for the follower outcome. No payload bytes were inspected. CC BY 4.0 in the Zenodo API covers the deposit, not linked software or model artifacts.

What we can build on

A useful communication study needs a defined signal, an observed receiver response and an independent outcome. Honeybees are a strong candidate for testing that chain, while existing claims and the single-site scope demand a narrow, independently reviewed question. This does not justify a general language model or claim that unknown animal speech has been translated.

The next reading task

Ask independent insect-behavior and ML/statistical reviewers whether any non-duplicative question remains and whether the relevant files support a bee-held-out evaluation. Any future follower-file intake needs a separately reviewed path that respects the 5 MiB per-file limit; do not bypass it with direct downloads, alternate tools or archive splitting. Keep fine-scale location metadata out of public outputs. No payload acquisition or model run is authorized by this note.

Go to the sources

Machine learning reveals waggle drift’s role in honeybee dance communication ↗
Primary study combining waggle-dance detection, individual tracking, dance followers and feeder observations in Apis mellifera.
Zenodo v1.0 honeybee waggle-drift data record ↗
Official versioned data record for the 2023 waggle-drift and follower-behavior study; its API lists 11 files and declares CC BY 4.0.

Original editorial note by Talk2Nature contributors. AI-assisted, source-grounded synthesis; not a peer review or an independent replication. Linked works retain their own rights.