An AI bot that teaches itself to walk, without a dataset and/or visual examples: that is Albert. And you can see how this works in this fun video.
In the video, which has been on YouTube for two weeks, you can see Albert. Albert is an AI robot that learns to walk. It does this by coming from five rooms, a kind of ‘escape rooms’. But there is something special about this process. In most cases, an AI robot learns to walk using a dataset and/or images of someone walking. In this case, the maker of Albert, whose name I don’t know by the way, chose to program it from scratch.
Deep Reinforcement Learning
Albert and the ‘learning environment’ was created through Deep Reinforcement Learning, a method of Machine Learning in which Albert is rewarded for performing actions correctly and punished for performing them incorrectly. For every attempt Albert makes, a score is calculated for how “good” it was and small, calculated adjustments are made to his brain to encourage behaviors that lead to a higher score.
Albert is controlled by an artificial brain (neural network) that consists of five layers. The first layer consists of the inputs (the information he receives before taking an action, such as the position and speed of his limbs) and the last layer tells him what actions to take. In the middle three layers, the calculations are performed to convert the inputs into actions. It is therefore updated after every attempt that Albert makes.
Go, AI Albert, go!
This is how Albert eventually learns to walk, overcome obstacles and challenges and turn. Be sure to watch the video, because it’s a funny and special process to see!