A new pick-and-place system developed by the Massachusetts Institute of Technology (MIT) is able to identify the best method of grabbing an object at lightning speed, without having prior knowledge of the object. The system can also classify objects so that they can be sorted into different bins, for example. In the future, the system can be used for organising products in a warehouse, among other things.
Pick-and-place systems are not new. Such systems are typically deployed in tightly controlled environments to perform one specific repetitive task, such as picking up a particular part of a production line. This operation is performed in exactly the same way every time, with each product presented to the robot in the same orientation. This makes such systems inflexible and unsuitable for picking up and sorting a large number of different objects interchangeably, for example.
Deep learning
The new system developed by researchers at US universities MIT and Princeton University does offer this capability. The system consists of a 'standard' industrial robotic arm, which the researchers have fitted with a custom-made gripper and suction cup. Using a deep neural network, the robot can identify how best to pick up an object, which attachment is best to use for this purpose and in which orientation picking it up is most likely to succeed. This is done without requiring knowledge of an object.
Such a deep neural network needs to be trained before it produces results. To make this possible, the researchers collected a large number of images showing bins filled with objects from the robot's perspective. For each image, the researchers marked which objects the robot could pick up and which it could not. These attempts were marked as a success or failure by the researchers. For each object, the researchers also indicated whether it is best picked up using the gripper or, on the contrary, the suction cup, and which orientation the tool should have in doing so.
This data was stored in a database, which was then used to train the deep neural network. Such a network is able to solve problems it encounters on its own by looking at previous similar situations with a positive outcome.
Identifying objects
Once the object is grabbed, the system lifts the object out of the bin. The object is then imaged from different angles using cameras. The images collected in the process are assessed using an algorithm, which links the image to stock photos stored in a database developed for this purpose. In this way, the robotic system can identify and classify unknown objects so that they can be sorted into different bins.
"This can be applied for sorting in a warehouse, as well as picking objects from a kitchen cupboard or clearing debris after an accident. There are many situations where picking technologies can have an impact," explains Alberto Rodriguez, professor of mechanical engineering at MIT and involved in developing the system.
Amazon Robotics Challenge
Rodriquez's team participated in July in the Amazon Robotics Challenge, an annual competition organised by e-commerce giant Amazon to promote innovations in warehouse technology. In it, the pick-and-place system was challenged to pick objects from a full bin or place them right in it. The robot was found to have a 54% success rate when picking up objects using its suction cup, while attempts to pick up objects with its gripper were successful in 75% of cases. The robot managed to identify all objects it was confronted with with 100% accuracy. In addition, the system was able to place all 20 objects in a bin in the time available.
Rodriquez was awarded an Amazon Research Award for his work and will further develop the pick-and-place technology in collaboration with Amazon. The focus here is on increasing the speed and responsiveness of the system.
Paper
The researchers will present a paper with more details about their pick-and-place system at the IEEE International Conference on Robotics and Automation in May.
Author: Wouter Hoeffnagel
Source: MIT
Source photo: screenshot of video
