The old robots can only do perform those functions what the programmers programmed within them. They could not perform in the unstructured environments where the situations are quite different. Hence they were unable to cope in the diverse environments. Thus neural network teaches robot new skills.
The MIT researchers have developed a new technique. The technique allows the latest robots to learn a new technique through handful of human demonstrations. It becomes able through machine learning methods. Hence the robot picks up the objects which are new to them and then places them at arbitrarily directions.
The machine learning model uses a neural network which actually recreate the shapes of the 3D objects. Through the use of neural networks, the machine actually learns all in just few demonstrations.
The availability of robotic arm and as well as the demonstrations are the only things needed. These things can effectively enable the robots to place the never seen objects such as mugs and bottles on their right place.
According to the researchers, teaching a robot have never been so easy. Their main contribution can effectively teaches the robot the new skill so that they can operate easily in an unstructured environment.
How Neural Network works
Without the use of neural networks, a robot can only be trained to perform only a specific item. For example, the robot can pick up the item from the left side only. If anyone places item on the other side then the robot sees it from a completely different scenario.
To overcome this issue, the researchers have developed a Neural Descriptor Field (NDF). NDF actually learns the 3D geometry of the items. NDF calculates the geometrical representation of a specific item by using a 3D point cloud. It coordinates in three dimensions. The NDF possesses the property known as equivariance. the robot is able to see the same object from any of its side. NDF is also also able to reconstruct the shapes of the same objects. The researchers can enable a robot to perform the task by simply pointing out its arm towards the object they want the robot to pick up.
The researchers have found the success rate of 85 percent. Success is entirely dependent on picking up a new object and then places it on its right direction. However, 2D image information is unable to cope with the requirements of the success criteria. But NDF has performed it so well.
Hence in this way, the use of neural networks can enable the robots to operate in the unstructured environments.