Robots come in all sizes and shapes, and they are increasingly common in workplaces, from factories to operating room. Many of the bots rely on attaining new skills by trial and error through machine learning. This new method allows such skills to be transferred between different robot shapes, eliminating the need to learn each task from scratch. “Practically, it’s important,” says Xingyu Liu, a computer scientist at Carnegie Mellon University and lead author of the research, presented this past summer at the International Conference on Machine Learning. “Research-wise, it’s a cool fundamental issue to study .”
,” says Xingyu Liu.
Let’s say that you have a robot arm and a human-like hand. You have taught it to use its five fingers to grab a hammer to pound a peg into the board. Now you need a two-fingered gripper for the same task. Scientists created a bridge between the two robots that gradually changes from their original form to the new. Each robot practiced the task and tweaked an artificial neural network until it reached a threshold success rate before the controller code was passed on to the next robot.
To transition between virtual source robots and target robots the team created a “kinematic tree”, a collection of nodes that represent limb parts, connected by links that represent joints. The team set the weights and sizes of the nodes for three fingers to zero to transfer hammer-whacking skills from the two-fingered gripper. The finger sizes and weights of each intermediate robot got smaller. The network that controlled them had to learn how to adjust. Researchers also modified their training method to ensure that the leaps between robots were not too large or too small.
The Carnegie Mellon system called REvolveR (for Robot Evolve-Robot) outperformed other baseline training methods like teaching the target robot from scratch. To reach a 90 percent success rate with the gripper, on the hammer task and in other experiments involving moving a ball and opening a door, the best alternative training method required from 29 to 108 percent more trials than REvolveR did, even though the alternative method used more informative training feedback. Further experiments were conducted by the researchers to test their method on other types of virtual bots. They also tried it with spiderlike bots to see if they could add leg sections to the gripper and have it learn how to crawl again.
” I think the idea was nice,” says Vitaly Kurin, University of Oxford computer scientist who studies robotics. Although it is not unusual to arrange challenges for an AI to transfer skills between tasks, he said, “This interpolation from one robotics robot to another one is something I haven’t thought of before .”