Humans aren’t the only ones working with Play-Doh. MIT CSAIL researchers have created a system, RoboCraft, that teaches robots how to work with the kid-friendly goo. The platform first takes the image of a shape (in this case, a letter of the alphabet) and reinterprets it as a cluster of interlocking particles. The bot then uses a physics-oriented neural network to predict how its two “fingers” can manipulate those spheres to match the desired outcome. A predictive algorithm helps the machine plan its actions.
The technology doesn’t require much time to produce usable results. It took just ten minutes of practice for an robot to perform roughly as well (and in some cases better) than humans remote-controlling the same hardware. That’s not the same as having a human shape the Play-Doh by hand, but it’s no mean feat for a machine discovering how to perform the task for the first time. Robots frequently struggle with soft objects where they tend to thrive with firm shapes.
RoboCraft-trained bots aren’t about to produce elaborate sculptures. The results are still imprecise, and the machine works slowly using just two fingers. The team is already developing a method of making dumplings, though, and plans to teach robots to use additional tools (such as a rolling pin) to prep the food.
The CSAIL scientists already have an idea of where the technology might be deployed. Kitchen robots could take over more responsibilities, while artistic automatons might create pottery. Eventually, technology like this could help the elderly and people with mobility issues by taking over household duties that require subtle motor skills.
Author: J. Fingas