German specialists from the Max Planck Institute for Intelligent Systems have created a robot dog Morty, who learned to walk using a neural network. In fact, it is used as a spinal cord and completely takes control. The scientists did not specifically program the ability to walk to enable the robot to do it on its own, but the goal of the study was to analyze and understand the whole process.
Initially, the neural network was not set up for perfect walking, so it had to adapt during the learning process. To do this, we used the Bayesian optimization algorithm, that is, optimization for an unknown function. The artificial intelligence imitated networks of neurons in the spinal cord that help animals walk by causing the muscles to contract in the right rhythm. In fact, the neural network created an instruction for the robot that it followed. He compared the information from the foot sensor with data from the virtual spinal cord.
At first, Morty fell and made awkward movements, then he adjusted and after about an hour he developed a smooth and stable gait. Scientists note that the virtual spinal cord has no idea about the limbs of the robot and its design, but can adapt its movements. In the future, they want to add additional sensors and expand the ability to move in space. The study should help understand how reflexes and learning work in animals.
Source: Trash Box