MIT Researchers Publish Massive Omnipush Dataset for Future "Pushy" Robot Projects
250 pushes of 250 custom-built modular objects gives researchers a wealth of data for future robots to mine.
Researchers at the Massachusetts Institute of Technology (MIT) have developed the world's largest dataset of how objects react when pushed — created by having a robotic arm physically push 250 specially-crafted oddly-shaped objects and recording the result.
Having a computer control a robot arm isn't exactly a new concept — but having the computer be able to figure out how objects will react when manipulated is considerably more challenging. A team of researchers at MIT is a step closer to the goal of a smart robot, though, thanks to an exhaustive dataset dubbed "Omnipush."
"We need a lot of rich data to make sure our robots can learn," says Maria Bauza, graduate student in the Department of Mechanical Engineering (MechE) and first author on the Omnipush paper, in an interview with MIT News. "Here, we’re collecting data from a real robotic system, [and] the objects are varied enough to capture the richness of the pushing phenomena. This is important to help robots understand how pushing works, and to translate that information to other similar objects in the real world."
Those varied objects, created specifically for the task, number an impressive 250 and were adjustable for shape, weight, and their centres of gravity. Each of these objects were pushed by a precision robot arm in 250 ways, with the result captured via a Vicon 3D motion-tracking system, depth-sensing camera, and traditional camera. The result: a data set which contains around 62,500 unique pushes and their outcomes.
Building on an earlier project which captured pushes on only 10 objects and didn't record video, the new dataset is considerably larger and more detailed. The modular objects were marked for accurate tracking down to a single millimetre, while the objects were modular and adjustable from 31 to 244 grammes of weight.
The results are impressive: the researchers claim that an algorithm trained on the Omnipush dataset proved twice as accurate at predicting the final position of pushed objects than the best of the competition.
The team's work is to be presented at the International Conference on Intelligent Robots and Systems; more information is available from the Omipush paper. The Omnipush dataset itself, meanwhile, can be downloaded from MIT MCube under a Creative Commons Attribution 4.0 International Licence.