KC Meat Up

Robots That Watch How Cattle Walk

Status: Proven (a stockman's eye and locomotion scoring) · Early commercial (camera lameness scoring, dairy only) · My own build (my robots filming gait) · Speculative (a pasture beef gait dataset and model)

A limping steer, a calf hanging at the back of the group, one with its head low and ears down: good cattlemen have caught these by eye for generations. I want my robots to watch the herd the same way, every day, and keep a record. Here's what exists, what doesn't, and how I plan to fill the gap on my small place of around 20 acres.

What a rancher's eye already does

The machines below are trying to copy that eye, not replace the judgment behind it.

What exists today

Technology What it does Maturity Published cost Fit for my small stocker herd (around 20 acres)
CattleEye (Belfast, UK) A 2D camera mounted at least 4 m up over the exit of a milking parlor. AI identifies each cow by coat pattern and scores mobility 0–100 every time she walks under it (Frontiers in Vet. Science 2023; Fortune 2025). Early commercial (dairy). On 3 farms and 6,040 scores, agreement with two vets was above 80%. For painful foot lesions it had 52% sensitivity and 81% specificity, vs. 29% and 89% for a trained vet (same Frontiers study). UK: £650 + VAT for a one-camera kit, then 50p/cow/month in year one and £1/cow/month after (CattleEye). Poor as-is. It depends on cows walking the same lane twice a day. Stockers on pasture don't have a milking parlor. The idea (a fixed camera at a spot cattle pass) is exactly what my water robot could create.
Research side-view and 2D/3D camera systems Single side-view cameras with deep learning (Scientific Reports 2024); fully automated 2D systems (J. Dairy Sci. 2025); 3D overhead video of hip height and spine curvature (UWE thesis). Research, almost all dairy. Wageningen's pose-estimation work on outdoor videos reached 80.1% classification accuracy using six gait traits; back posture, head bobbing and tracking distance mattered most (Russello et al. 2024). Not sold. The method transfers. The setting (a fixed walkway) doesn't, yet.
Beef cattle, 3D camera University of Illinois filmed 189 Angus × Simmental steers with an Intel RealSense depth camera and measured step length and hind-leg angle. Step length explained 64% of the variance in average locomotion score (Dawson et al., U. of Illinois). Research, indoor slatted barns. Off-the-shelf depth camera; no system for sale. Closest beef evidence I found. Step length is something a camera on my robot could measure.
Accelerometers (ear tags, collars, leg sensors) Measure activity, lying, rumination. In 508 weaned dairy calves, an ear-tag accelerometer flagged respiratory disease 2–3 days before clinical diagnosis at 71.4% sensitivity and 95.2% specificity (Animals 2022). On pasture, accelerometer + GPS behavior separated lameness scores in dairy cows (Prev. Vet. Med. 2021). Early commercial for activity and rumination; lameness from sensors is still research. Varies by vendor; none I can cite for beef stockers. My Nofence collars already carry GPS. Movement data plus video is a stronger pair than either alone.
Drones Herding and behavior research (Applied Animal Behaviour Science 2023). I found no peer-reviewed study of drone video scoring lameness in grazing cattle. n/a Speculative for gait. Useful today for counting and finding a straggler in a far corner.
BRD and stragglers Machine learning on behavior and sensor data for early BRD detection is an active research area (Frontiers in Animal Science 2022). Research to early commercial (mostly feedlots and calf barns). n/a The "who's at the back, who skipped water" signal is what I care about most for new stockers.

Public datasets I can learn from

All are dairy, mostly a cow walking a lane:

What doesn't exist yet

As far as I can find:

You're right that it should exist. The pieces do: cheap cameras, pose estimation, collars, and a scoring scale cattlemen already trust.

My plan: the water robot as the camera lane (My own build)

Cattle walk to water every day. If the water moves, the walk to it moves too, which turns my water robot into a roaming version of CattleEye's parlor lane.

  1. Camera on the trough robot. A side-view camera (and later a depth camera like the Illinois study used) aimed at the approach path. Cattle are filmed walking in, every day, without being gathered.
  2. Who is it? EID tag readers at the trough plus the Nofence collar position to tie each clip to an animal.
  3. Label the old way. I score each animal 0–3 on the Zinpro scale on my own walks, and my vet confirms any animal treated. Pairwise "who walks worse" comparisons help where I'm unsure, following the UBC method.
  4. Track the stragglers. Collar GPS tells me who's last to water, who's apart from the group, and who stopped moving. That's a flag to go look, not a diagnosis.
  5. Measure first, model later. Step length, head bob and back arch from pose estimation, compared against my scores.

What I'll measure: clips per animal per day, how often video and my eye agree, and how many days earlier (if any) the robot flags an animal than I do.

The dataset (Speculative)

With 6–8 head, my first season gives a small dataset. Small herds won't build a model alone. The Speculative part is a shared, open dataset of beef cattle walking on pasture, contributed by many small operations, labeled the same way, published with consent and with each producer's data under their control. I'd publish my clips and scores on these pages, starting with the research page.

Where the eye still wins

A camera sees a limp. A cattleman knows the steer came off a rough truck, that it's foot rot season after a wet week, or that the calf limping is the one that always walks oddly. The robot's job is to make sure I look at the right animal on the right day. The judgment stays with me and my vet.

Sources

Full links are on the research page.

Helping Kansas City get healthy with meat — powered by agents.