Last September, my boyfriend and I adopted a 7-month-old, 45 pound husky mix from an animal shelter here in New York City. But before we even brought our new pup home, we hired a trainer. I grew up with badly behaved dogs—good dogs, but poorly trained ones. And I knew that if we were going to be a happy family we would need to make sure to teach our new pooch what her role was, and what we wanted her to do (and, perhaps more importantly, not do).
Fast forward eight months and Moro is a really great dog. She’s extremely friendly, rarely barks, doesn’t feel the need to guard people or food or toys, and loves to just hang out on the couch with me. She’s sitting at my feet while I write this. But we’ve worked hard to get her here, and spent a lot of time and money training her.
She used to pull like a sled dog on the leash, and when she would see another dog (regardless of whether that dog seemed excited to see her) she would get so excited that she’d leap around and yodel in delight, scaring everybody around her. She used to wail and whine when she wanted something, and scream if she was separated from me.
Today, she walks right by my side, runs next to me every morning, and ignores other dogs unless I release her to say hi. No leaping, yodeling, or pulling. She asks for permission before coming onto the couch or bed. She comes when called, and stays when she’s asked to. We’ve still got some work to do of course—squirrels are still a wild card and every so often she’ll catch a scent that’s just too interesting to ignore. But she doesn’t chase cats in our yard, and most of the time will leave the bushy tailed rodents alone unless they really tease her.
This is only partially a brag, but I also feel like I deserve to. I’ve worked hard to turn Moro into a really well behaved dog. I spend hours every week doing training drills with her, teaching her new things, practicing our manners, seeking out new situations to test her in. We’ve hired several trainers, and even sent her to a two-week doggy training bootcamp while we were traveling. Some dogs are naturally well behaved, but most need to be taught what we expect of them, and dog training is hard.
So when I saw that a research team was working on an algorithm to train dogs (or, in science speak, a “a Computer-Assisted Canine Posture Training System”) I was instantly intrigued. Could I outsource all this time and money to a computer? Are robot-doggy trainers in our future?
The short answer, of course, is no. But Dave Roberts and Alper Bokzkurt at North Carolina State University are trying to help amateur trainers like me become more efficient. And they’re trying to do that using a series of sensors and algorithms.
Roberts and his team at NC State University are interested in the ways that sensors and algorithms can help humans and dogs communicate better. Last year, they announced a project that helps handlers monitor guide dogs using a harness and vibrating handle. And one of the things they’ve noticed about trainers like myself, is that we struggle with giving consistently timed feedback to our dogs.
Professional trainers have a lot of experience making sure that command and reward are well timed: the dog gets the reward quickly enough after following the command that she associates the two, but not so quickly that she doesn’t finish the behavior. (Moro likes to hover her butt just off the ground in an attempt to defy our requests that she sit.) Humans are just not very good at timing things exactly. But computer are. So Roberts is working on an algorithm that rewards the dog at exactly the right time, every time.
The system in question in this case is relatively simple: each dog wears a harness that has a series of sensors on it in the front and the back. The sensors can pick up whether the dog is moving and where its different parts are, and they send that data to an algorithm that can figure out what position the dog is in: sitting, standing or lying down.
To train the algorithm, the dog is fitted with one of these harnesses and brought into a little pen. The owner of the dog tells it to sit, and when it does, the algorithm crunches the data coming in from the harness. Once it decides the dog is sitting, it sends a signal to an automatic treat dispenser in the pen. The dog sits, the computer recognizes the sit, the dog gets a treat. And repeat.
You can see the process in the video below, in which Simba very quickly learns that sitting means tasty treats. (Simba is a guide dog for a member of Roberts’s lab and is often used in their experiments as a test subject, so he is already exceptionally well trained and knows how to sit. The point here wasn’t to train him to sit, but rather to train the algorithm to recognize sitting.)
The point here, Roberts reminds me several times when we talk, is not to get rid of human trainers. “We will never get rid of humans in this process. The human animal bond is really important to all the things we do,” he says. “But just like technology gives humans new ways to interact with each other, it can also give humans and animals new ways to communicate.” What he hopes to do is combine this algorithm with the humans who are trying to do a better job training their pets, to help them get better at timing.
“Maybe if you have an Apple watch or a smartphone that is getting signals from the sensor that we’ve designed and will alert you and say ‘hey you need to perform this action, distract your dog, cue your dog, get your dog into a sit, stand in front of your dog,’ If computers can help novices make these decisions more effectively, then it’s easier to do these behaviors.”
So Roberts doesn’t think he’ll ever replace trainers, and neither does Katenna Jones, the certified dog trainer I asked about the system. “Currently, I'm skeptical,” she wrote in an email. First, not all dogs are comfortable wearing a vest, she says. “If you throw a vest on an adult, fearful dog and a machine talks to it, it can backfire. Many dogs will be fine with it, but many won't.”
And right now, the system only teaches sit and down. Which are the easiest things to teach a dog to do. “If a trainer can't teach sit and down, they're not a trainer,” Jones said. She has a whole list of things dogs should be able to do aside from sit and lay down: “Reliability amid distraction, without the reward, in different locations, for different people, etc. I would say just leaning sit and down in the home are barely training.”
Problems like the ones we faced with Moro: overexcitement around other dogs, cat chasing, ignoring distractions, pulling, focusing on me, those are all really hard to program an algorithm to train. And those are the things that really make a dog reliable and well behaved.
But even if there was a super-advanced robot that could train your dog for you, it wouldn’t train your dog to listen to you. “It's very common for dogs to listen to one person in the house, but not the others. The same would absolutely apply here,” Jones said.
“The human learning how to do it is the most valuable part: the bond, the communication, the excitement at success. They can then carry on the concepts with other skills, beyond sit and down. Otherwise I could just walk into a home and train dogs. When the owner comes home, voila, the dog is trained. It just doesn't work that way. If it did, my job would be easier and I'd have a lot more money!”
But she sees promise here too. “I see tons of potential for dogs residing in shelters or laboratories who don't get enough mental and physical stimulation,” she told me. “Adopting a dog that sits and lies down is very appealing (assuming shelters can afford such an expense).”
Studies have shown that if a dog knows how to sit on command, they are more likely to be adopted. And one of the main reasons that animals are returned to shelters after being adopted is “behavior problems.” If a shelter can’t afford to hire more people, but can afford a system that helps train their dogs, it might be worth it to save more animals. Roberts mentioned this to me too, saying that a system like this could help time-strapped volunteers train shelter dogs more efficiently.
One thing Jones worries about with all of this dog training technology, is that it’s sending the wrong message to dog owners. “I think humans are relying on ‘quick fixes’ far too much. Automatic feeders, doggie doors, now robot trainers. If you don't have the time to put into training, you shouldn't have a dog. If you're using this tool with a trainer, or it's teaching the human, or it's just a fun toy—sure! But if it's to replace training? No way.”
And before you write her off as simply worried about being replaced by a robot, she’s not. “I'm not concerned about job security. I'd love to see a day when trainers aren't needed because people know enough about dogs. But that day will never come. In fact, the need is growing.”
In the press release and paper, the team emphasizes the ways this system could be used to train working dogs, dogs that do things like search and rescue, bomb sniffing, or assisting the disabled. But Roberts admits that that’s not why he’s doing this research.
“My heart is really in the pet space,” he says. According to the ASPCA, every year 1.2 million dogs are euthanized in shelters across the United States. “I don’t think it’s fair that our limitations and humans is the cause of them to be put in shelters and euthanized.” Roberts hopes that this system might be able to help people address problem behaviors that land dogs in shelters, or help shelters get dogs adopted more easily. “I’ll be honest, my heart is there,” he says.
As for Moro, we’ll stick to our current training methods and leave the algorithms to the researchers for now. Unless the computer can teach her not to fart on my pillow, in which case, sign me up.
