Robert Gleichman Rose Hills
Revolutionizing Robotic Manipulation: Learning to Fold Towels from Demonstration
Currently, it is difficult for robots to manipulate deformable objects. This is because deformable object’s plethora of states make it is difficult for the robot to figure out what next action will get it closer to its goal. One solution to this problem is to use human demonstrations. The robot can currently tie knots when given a demonstration of a knot tying sequence. This is done by finding a transformation from the rope state at test time to the rope state at the beginning of a demonstration. This transformation is then applied to the motion of the robot’s grippers recorded during the demonstration to create a motion of the arms that will work for the new rope state. This summer, I hope to extend upon this capability by allowing the robot to choose the best out of many demonstrations, and teaching the robot how to fold towels and garments.