This team of UC Berkeley researchers has developed algorithms that enable their PR2 robot, nicknamed BRETT for Berkeley Robot for the Elimination of Tedious Tasks, to learn new tasks through trial and error.
This article could have been much better if it started with this: Deep learning is not new. Using deep learning on motor tasks is, but I guess that doesn't make for a thrilling headline.Applying deep reinforcement learning to motor tasks has been far more challenging, however, since the task goes beyond the passive recognition of images and sounds.
This end-to-end training process underlies the robot's ability to learn on its own. As the PR2 moves its joints and manipulates objects, the algorithm calculates good values for the 92,000 parameters of the neural net it needs to learn. I have to wonder what the human equivalent would be, if that analogy can be drawn. How many parameters must be established for a human to throw a baseball? It will be a strange flip when robots are seen as quick, agile, and adaptive, and humans as slow, clumsy, and rigid.BRETT takes in the scene, including the position of its own arms and hands, as viewed by the camera. The algorithm provides real-time feedback via the score based upon the robot's movements. Movements that bring the robot closer to completing the task will score higher than those that do not. The score feeds back through the neural net, so the robot can learn which movements are better for the task at hand.