Researchers from the University of Washington put a GoPro video camera on the head of the Alaskan Malamute, and position sensors on the legs and tail. In this outfit, the dog took part in the recording of 380 short clips demonstrating the animal's behavior in various situations. These data formed the basis of the training program for the new neural network DECADE.
50 different interiors were used for recording: living room, office, park, street, etc. The dog was given simple, most common commands, played, walked and interacted with several categories of people, other dogs and machines. In each lesson, the position of the limbs in space, the dynamics of movement, sounds and visual images were recorded so that the AI could draw connections between them.
The main task for the neural network at this stage is to answer the question "what the dog will do in a particular situation." For example, it will wave its tail at the line "Good boy", but will remain in place. And if you throw the ball, it will run, but how exactly will it bend around obstacles? DECADEs demonstrate the beginning of the scene and the neural network must simulate its continuation.
The ultimate goal of the project is to prepare a platform for constructing robotic dogs, or machines that imitate other animals. Evolution has already taught them to find a safe route in the forest, to effectively pursue prey or to escape, to defend themselves. And instead of coming up with algorithms for robots from scratch, isn't it easier to just learn the basics of their behavior from real animals? In the next version, AI will be trained to work with smell and touch.