If you were buying a kitchen knife in a supermarket, you wouldn't expect the cashier to swing it dangerously close to you as they were ringing it up. If that cashier were a robot, though, it wouldn't know any better – unless it had been taught otherwise. That's just what engineers at Cornell University have done, using a unique new technique.
The scientists, led by Prof. Ashutosh Saxena, were working with a Baxter industrial robot manufactured by Boston's Rethink Robotics. That particular robot has recently received a lot of good press, not only because it's relatively inexpensive, but also because ordinary human workers can teach it new tasks just by manually guiding its arms through the required motions.
Cornell's Baxter was set up in a mock grocery store checkout, where it visually identified different types of goods, then proceeded to pick them up and transfer them to a bagging area. Initially, although it was able to sense the proximity of nearby human bodies via its heat sensors, it didn't "care" if it passed a knife perilously close to them.
To train that behavior out of it, participants started by reaching in and stopping the robot in mid-swing, pushing its arm back to a safer distance. The next time around, the robot used its screen to display three possible better trajectories for its arm-swing, based on what it had just learned. Participants would select the best one.
On the robot's third try, the participants once again reached in to guide its arm, this time fine-tuning its movements, and instructing it to swing its "hand" around so that the knife's blade wasn't facing the customer. In this way, using custom learning algorithms, the robot was ably to learn incrementally how best to handle knives, with each lesson drawing upon what had been taught in the one previous.
Three to five lessons were all it took to teach the robot how to move a knife. The process could also be used to teach it how to move a variety of other items, such as easily-squished fruits, or packages that need to be kept upright.
More information is available in the video below.
Source: Cornell University