Boston Dynamics' Spot the dog has come a long way since the four-legged robot was invented, and now it looks like the old-timer has learned a new trick: how to play fetch.
It's certainly not the most dynamic of fetch games and springs images to the forefront of your mind of an old dog creakily picking up a ball and dropping it in front of its owner in slow motion. But it's a new trick, nonetheless. You just have to imagine the mechanical tail waving from Spot, as he anticipates the next throw of the ball.
Boston Dynamics is fully aware that Spot wasn't built to replace man's best friend, and doesn't expect everyone to dish out $74,500 for a game of robotic fetch in the park. However, as the company stated in its blog, this new development could serve as the basis of future, more practical uses, like picking up litter on the side of the road.
In some way, it's nice to see that Spot's life doesn't have to revolve around all work and no play. In other playful instances, Spot has been taught to "pee beer," and to dance to music, while on the working field it's been taught how to use its arm attachment for useful factory work, and how to assist with data collection in nuclear power plants, among other uses.
Now, Spot can play fetch. But how?
How Boston Dynamics trained Spot to fetch
Generously, Boston Dynamics has shared a tutorial on how it taught Spot to play fetch, so that anyone who wishes to train their robot dog can do so too.
How do you teach a robot dog new tricks? Spot's API makes it simple to create new behaviors like this machine learning-inspired game of fetch. Read the blog to see how it works. https://t.co/FJoBEfrYIL pic.twitter.com/KjZ7835gX8— Boston Dynamics (@BostonDynamics) June 4, 2021
In the tutorial, you can learn how to create a fully functional API example. In doing so, you can train a machine learning model to detect a toy, how to command Spot to pick it up with its arm, how to help Spot detect a person, and how to drop the toy two meters away from that person. Just like you train a dog to play fetch, just with computers.
The most time-consuming part of the process, perhaps, is the creation of a large database of images captured by Spot's camera, which have to be manually labeled, and which are then used for the machine learning model.
There have to be approximately 400 images snapped for each environment, so if you first train Spot to play fetch in your garden, you then have to start the entire process from scratch if you then play on the street or in the house.
It sounds time-consuming for a game of not-so-thrilling fetch, but if you have time on your hands, it's worth a shot.