BEHAVIORAL DATA SCIENCERESEARCH LAB
BEHAVIORAL DATA SCIENCERESEARCH LAB
A live look at a single planarian flatworm in its dish. The stream below is the raw rig; the panel underneath is its position and movement, tracked automatically in near-real-time (location, speed, distance from center, and turning) and refreshed continuously.
No sign-in needed. A blank player almost always means the stream is just briefly offline — the rig reconnects on its own, so wait a minute and reload, or open it directly on YouTube. If it's still blank, a privacy or ad-blocking extension may be blocking the embed — pause it for this site and reload.
The video you see above is real-time. The tracking plot takes a different path: 1-minute clips upload to Google Drive, and a second Mac pulls and processes them with a YOLO tracker + behavior model every 90 seconds, then publishes the plot here - so the data trails the video by about 3 minutes.
YOLO detector + temporal cleanup, scored against 200 hand-labeled frames.
For scale, the dish is about 50 mm across — a typical error is roughly 1% of the arena.
5-class model, leave-one-out cross-validation (F1 per behavior; 1.0 is perfect).
Numbers are from held-out evaluation on this rig's footage. On the live plot, the locomotion states (gliding / resting / turning) are read directly from the worm's motion, while the model supplies the postural calls (contracted, wig-wag).
None of this is automatic magic — every model here learns from frames and clips a person reviewed by hand, one at a time. The running totals:
Hand-labeled on this rig's footage.
Labeling is the bottleneck, and it's something anyone can do — no biology degree required. Mark where the worm is, or tell us what it's doing, and you directly improve the models running on this page. Add your name to join the contributors.
Label where the worm is → Label what the worm is doing →