Robot grabbing is successful in 98% of cases. Training a robot how to grasp various objects without falling usually requires a lot of practice.

 

However, researchers from the University of California, Berkeley and Siemens jointly designed and described in an upcoming paper a new type of robot that can learn how to grasp new objects by studying a database of 3D shapes. The robot is connected to a 3D sensor and a deep learning neural network. Researchers use these two to provide image information of the object. This information includes object shapes, visual appearance, and physical knowledge of how to grasp them.

 

Therefore, when a new object is placed in front of the robot, the latter only needs to match the object with a similar object in the database. In actual operation, when the robot is more than 50% confident that it can grasp a new object, it will succeed in 98% of the cases. However, if the robot's confidence is less than 50%, it will first tentatively grasp the object, and then form a grasping strategy. In this case, the robot has a 99% chance of success. So the way to overcome the lack of self-confidence of the robot is to do a quick little check.

 

This training method can reduce a lot of machine learning time and make the robot more flexible. Jeff Mahler, a postdoctoral researcher working on this project, told MIT Technology Review, "We can generate enough training data for deep neural networks in one day, eliminating the need to run several months of physical experiments on a real robot. The trouble.” The robots currently in use in factories are already very accurate in grasping known objects, but they still can’t adapt well when facing new objects. The efficiency of this training strategy and the reliability of the robot's grip strength enable this method to play a good role in future commercial applications.