This AI Thinks Like a Dog

An AI-powered robotic dog? That’s amazing because we are talking about a machine behaving exactly like a biological being. For artificial intelligence enthusiasts, we call this great progress towards a future where agents will serve and accompany us, throughout our lives.

You’ll not have to worry about medical issues on your artificial dog leave alone losing it to death, which is a great idea! Like 20 years ago, a neighbor poisoned my dog, and I still remember it to date, wish it were robotic, like the one currently being created that acts and behaves exactly like a dog and makes canine decisions.

Nobody Wants to Think of Their Dog Dead

Do you? The loyalty of this animal friend is unbeatable. In fact, some people can’t live without a dog. The blind need it for guidance on their way while cops use it to trace lost individuals as well as sniffing out forbidden drugs. On the same note, it’s clear that not everybody can afford to buy even a puppy because the creature can be super expensive. That’s why experts take it as a brilliant idea to create an AI-powered canine that can handle such tasks as well.

That is, Kiana Ehsani and his team from University of Washington, Seattle have managed to successfully train an AI system; to act, think, and make decisions similar to those of a real dog. Well, it may sound crazy but don’t forget about the irreplaceable value of a biological dog even after it’s dead (thank goodness at least the artificial one won’t die.)

The Project

Source: MIT

The team gathered a special dataset from the dog family, which in essence brings up the possibility of replicating the same idea on other intelligent creatures we interact with. Ideally, building a database around canine behavior may not have been something tedious because often times, the animal is cooperative. The researchers singled out one dog, took measurements of its head, body, tail, and legs, captured and recorded the angles at which the parts turn in different situations and listed why.

In addition, a special GoPro camera was set on the animal’s forehead to capture the visual scene, and the data stored after every 5 seconds. To collect voice data the scientists placed a microphone strategically on the dogs back near the neck region – with the sound saved on the Arduino unit.

In exact figures, the researchers collected a total of 24,500 video clips which coincided with various body movements, positions and the situations the animal was in. They used three-quarters of the data to train the AI system and the rest to validate and test its capability.

The Practical Goal

Like training a dog, the major purpose of the project was to help the system learn progressively on its own in future. That is, how to act and plan like a dog — with long intention being, how to learn from biological dogs in the real world.

The first test was to have the AI predict the expected next move the dog is likely to make, with reference to the images presented. The system manages this by analyzing the videos and learning from what the natural dog did initially.

After observing five images in sequence the system was able to smartly tell the next moves, which Ehsani says, “Our model can precisely predict the moves of the future dog just by looking at the actions illustrated in the images captured.”

Next was the planning aspect, which proved quite tricky because it involved referring to the data in various angles to find a special pattern of actions, and the reason that made the dog move to the location it did. This was also achieved by presenting a wider collection of the sequential videos recorded.

Now, the interesting part came, where the system again managed to learn where dogs can and can’t stand, or walk, which ideally equals to mastering doggish behavior. So the engineers used the dataset to teach the AI about diverse kinds of surfaces and how to identify them at other locations different from what’s seen in the videos. “With ego-centric videos, our model has learned to plan, move and act like a dog in a similar situation,” the team said.

While there is still plenty of research and implementations to be done, at least it’s clear that the future dog being developed will have an almost near natural brain looking at the current progress of the model.