If Wall-E Was a Farmer: Teaching a Robot to Herd Cows

By Rachel Swaby

How researchers from the Australian Centre for Field Robotics spent months considering how to modify a general purpose robot for interaction with the slow-moving livestock. Wall-E, meet cow.

What was most surprising about the cows last April in Camden, Australia was that they didn’t seem to care that what could have been a distant cousin to the Mars rover had strolled in from the lab and claimed authority in the paddock like it was the 4th rock from the sun. Apparently without allegiance to the humans who gave two hours a day moving them toward the milk barn or to the dogs who were cow-wranglers by trade, the robot that assumed the herding duties was simply accepted as a regular fixture in the daily routine of foraging, ambling, and evacuating milk.

The cows’ blasé response was the best possible outcome for University of Sydney researchers from both the Australian Centre for Field Robotics and the department of Veterinary Science, who had spent months considering how to mod the general purpose bot for interaction with the slow-moving livestock. Rounding up cows to for milking isn’t a particularly difficult chore for farmers, but it occupies an hour in the morning and another in the afternoon during an already jam-packed day. And in Australia, the task is often carried out on a quad bike, which is one of the leading causes of injury on a farm.

Kendra Kerrisk, an associate professor of veterinary science, whose work focuses the future of dairy farming, identified the cow-wrangling as ripe for automation: “In winter, it’s freezing and in summer it’s really hot and dusty. Farmers try to do the milking as quickly as possible and get the cows home faster than they’d go on their own, which is not a good situation.” See, a cow’s impressive 330-degree vision has a blind spot: the terrain just about to be explored with their hoof. Herded too quickly, and missteps could lead to stone bruises and lameness.

A day’s most admired quality, according to a heifer, is predictability. If programed to herd at a slow, consistent pace, thought Kerrisk, a robot could give cows the time they need to change locations, while also freeing up man-hours. So Kerrisk asked the Australian Centre for Field Robotics if they’d help her use their “perception research platform” for a test run.

Although the robot had been in agricultural service before, its previous post was in the orchard, surveying fruit trees to make judgements about ripeness and disease. Apart from getting one of its four wheels stuck in lumpy terrain or misjudging its proximity to a tree trunk (both unlikely), it was pretty low risk work. Working with cows would introduce new challenges.

Cows are enormous, delicate animals, susceptible to any number of adverse conditions, none the least of which looks like Wall-E on steroids.

Cows are enormous (upwards of 1000 pounds), delicate animals, susceptible to any number of adverse conditions, none the least of which looks like Wall-E on steroids. So if there was some hand-wringing by researchers in advance, it was in considering how the research facility’s cows would take the robotic disruption. It was important that no harm come to either. To protect both animal and machine, the researchers outfitted the robot with a metal skirt—a sort of bumper that a cow could scratch its back against should it mistake the robot for a shrub. The skirt also enforced some distance between the robot’s expensive cameras and the cameras’ living subjects.

For a machine built in 2009 with urban environments in mind, the transition from studying the movement of people and cars to interacting with grazing cattle was an easy one. The robot already came equipped with many ideal traits for the field. “The vehicles are much lighter than typical farm equipment, which is beneficial for reducing soil compaction, and they are electric powered, meaning they are quiet yet powerful,” says James Underwood, a senior research fellow at the Australian Centre for Field Robotics at the University of Sydney.

When the bot was deposited in the farm in April, Underwood explains it was wearing all its best: “colour stereo video cameras similar to a hand held camera, yet also providing depth sensing like human eyes; thermal cameras that can remotely measure heat; 2D and 3D laser and RADAR ranging sensors to build real-time 3D models of the world; navigation sensors to help locate the vehicle; hyperspectral sensors to measure light in more sophisticated ways than regular cameras; and also soil radiation and conductivity sensors.” Although autonomy is the ultimate goal, its first meeting with the cows was arranged via remote control.

So as not to spook the subjects, the 4-wheeled robot first tiptoed along the outside of the fence line in 10-minute intervals, offering the cows an opportunity to observe without entering their physical space. As it moved, scientists took a tally of how many cows approached the robot and how many fled. By the third lap, it proved to be an unremarkable show and the robot was ignored.

Upon strolling into the paddock, the robot performed a big figure-eight to make its presence known.

What happens when a robot strolls into a paddock? Remarkably little. The robot performed a big figure-eight to make its presence known. The first time it approached a cow, the animal would scurry away, but on subsequent loops, the cattle held their ground longer. They were rapidly desensitized the the machine in their midst.

The swoops and scurries were observed by researchers on a monitor from two football fields away. The robot’s operator was steering it based on the 3D real time movements of the cows streaming from the mobile machine to the research station. “It was similar to playing a video game on a laptop,” said Underwood. “It was actually quite fun!”

When it came to rounding up the cows, the robot excelled, herding 20 cattle out of the paddock on three tests. The robot was also able to herd a drove of 150 cows on two occasions! The cows made their journey at a comfortable and consistent clip, without a stone bruise incident or too much hemming and hawing.

The data gathered during the test—how the cows move in a variety of situations—will contribute to furthering the research. “There was a lot of promise for automating the test,” says Kerrisk. The data collected will be used to design algorithms that dictate how the robot should respond to a straying cow or a cooperative group. The result could be a robot rounding up the cows all on its own. “We have a large amount of data for the initial dairy herding application, so the next step is to work out the opposite question: how little data do we need?” says Underwood. “Then we can design a next stage prototype specifically for this industry, at a much lower cost, while having the confidence that what we’re trying to do is achievable.”

When Kerrisk presented her research, farmers were excited. But both Kerrisk and Underwood believe that automating herding is just the first step in a long line of possible innovations. “Well, if you’re going to have a robot bringing the cows home, it becomes a very expensive dog,” says Kerrisk. “Some will be happy to pay for an expensive dog and some will not.” The real value, says Kerrisk, are all the possible add-ons. What if the robot could tackle other farm issues, too, like spotting fences that are no longer electrified or monitoring the moisture or other environmental conditions in the soil? The robot could also figure out how well the cows have grazed the paddock and alert farmers if there seems to be a problem with a cow giving birth. The researchers think herding is just the baseline. “Our next prototype will be even tougher and more capable. We are in the process of putting together a next-generation platform that is even better suited to the often rugged terrain found in typical dairy environments,” says Underwood.

Since leaving the dairy, the robot had traveled north to a tropical region of Australia, says Underwood, “gathering data on banana, mango, lychee, custard apple and avocado farms. A robot as flexible as this finds itself on a tight schedule.”

Photos courtesy Australian Centre for Field Robotics