Unlike other AI devices, the robot doesn’t require human-provided models and simulators or a long trial-and-error process to operate on its own. “If we want robots to become independent, to adapt quickly to scenarios unforeseen by their creators, then it’s essential that they learn to simulate themselves,” Hod Lipson, a Columbia University mechanical engineering professor, said in the press release.
For the research, Lipson and his Ph.D. student Robert Kwiatkowski tested a four-degree-of-freedom articulated robotic arm. In the beginning, the robot moved randomly and collected roughly 1,000 trajectories, each compromising 100 positional data points. Then, the robot used this deep learning information to create a model of itself. These first self-models weren’t on point though: The robot didn’t know how its joints were connected.
However, after less than 35 hours of training, the robot produced a self-image that was accurate to within approximately four centimeters. Next, the self-model conducted a pick-and-place task in a closed loop system that allowed the robot to pick up objects and deposit them into a receptacle with 100 percent accuracy. With the closed loop system and an internal self-image, the robot completed the pick-and-place task with a 44 percent success rate.
[
link to www.geek.com (secure)]