The new RoboCraft device can squeeze play dough, release it and make different letters out of it.
"Modeling and manipulating objects isan important skill for robots so they can quickly learn to perform complex industrial and household tasks such as stuffing dumplings, rolling sushi, and making ceramics,” says Yongzhu Li, a CSAIL graduate student and author of a new paper on RoboCraft.
However, the authors note that objects withhigh plasticity, such as dough or plasticine, are not well understood. Therefore, with the help of RoboCraft, the team learns dynamic models based on multidimensional sensory data.
When the robot uses smooth materialindefinite form, the entire structure must be taken into account before any simulation is performed on the object. The authors turned the images into graphs of small particles, used a graphical neural network as a dynamic model to make more accurate predictions about the change in the shape of the material.
The first part, perception, isto learn to see an object (plasticine). The robot uses cameras to collect visual sensory data from the environment, which is then turned into small clouds of particles that can be used to infer the shape of an object. The graph-based neural network then uses the specified particle data to model the object's dynamics and how it moves. After that, the algorithms plan the behavior of the robot.
Now the team wants to teach the robot how to wrap ready-made fillings in dumplings.
Read more:
Scientists uncover how vitamins affect the incidence of cancer
The space probe flew 200 km from Mercury. Look what he saw
The satellite of Jupiter was looked at in a new light: what scientists saw there