A robot can design, build and test objects made from folded paper, such as planes or grippers, better than a human if given the same number of attempts.
Robotic laboratories can test and design materials far faster than humans, but they often rely on computer simulations to cut down on real-world testing for the robot. However, this doesn鈥檛 work when testing objects that are difficult and computationally expensive to simulate, such as fluids or deformable materials like paper.
Now, at Columbia University in New York and his colleagues have developed a robotic testing platform, called PaperBot, that can design objects made from paper, without needing computer simulations. 鈥淲e wanted to design tools in the physical world directly, instead of in the simulation, because in this way we can model many more realistic behaviours that are hard to simulate,鈥� says Liu.
Advertisement
To make a plane, the robot is first given a rough outline for a folded paper design, but is allowed to vary the length and width of the wings. The robot then folds the plane, chooses a launch angle and throws it. After measuring how far it flies, the robot adjusts the design using a machine learning algorithm and tries again.
鈥淧aperBot works very similarly to how humans do things,鈥� says Liu. 鈥淲e try random, different designs and then our memories remember what the good ones and the bad ones look like, and we try to find a pattern there.鈥�
After 100 trials, which took about 3 hours, PaperBot鈥檚 best plane design flew further than the best plane designed by a person in the same number of attempts at optimising the wing design before letting the robot arm throw it.
In a separate exercise, PaperBot was tasked with optimising a robotic gripper made from paper using kirigami, the Japanese art of paper cutting. It found a design that could lift objects weighing around 100 grams, the equivalent of four strawberries.
arXiv
