Machine learning benefits for industrial and collaborative robots
Machine learning is advancing the capabilities of collaborative and industrial robots. Without 3-D sensors or neural networks, robots are blind and one-dimensional. They’re restricted to one repetitive task that’s been preprogrammed with no ability to account for variables in their environment. This limits a robot’s productivity potential. Now, with vision sensors and machine learning capabilities, collaborative and industrial robots are able to achieve far more than they ever could on their own.
Robot and machine learning possibilities
A recent application of machine learning in robotics comes from UC Berkeley and Siemens with their DexNet 2.0 robotic system, developed last year, to pick up parts that it had never seen before. Training a robot to grasp objects without dropping them requires quite a bit of programming, practice, and trial and error.
This robotic system, leveraging a 3-D sensor and deep-learning neural network which processes information on the shape and appearance of an object, as well as how to grab it. The robotic system is 98% accurate when it is at least 50% confident it could grab a new object. If it was less than 50% confident, the system would perform a quick visual inspection, and then grab the part with 99% accuracy.
This capability for robots could transform the way material handling robots are deployed and programmed in commercial applications.
Types of machine learning in robotics
There are different types of machine learning in industrial and collaborative robotics. The example above is an advanced version of computer vision or robot vision. Essentially, complex optical equipment for image capture feeds neural networks so that a robot can "see." In most instances, this translates into robotic guidance to avoid collision, seam tracking during welding, and to ensure parts are grasped correctly.
Another fascinating new type of machine learning in robotics is imitation learning. Essentially, in this scenario a robot can be programmed by demonstrating how to complete a task. For example, someone could show a collaborative robot how to grasp an object by guiding the robotic arm the first few times. In this way, the robot would learn to grasp the object on its own.
There are other types of machine learning in robotics, such as self-supervised learning or multi-agent learning, but imitation learning and computer vision are two of the main methods.
Machine learning opens up entirely new possibilities for industrial and collaborative robot applications, allowing both types of robots to perform tasks that were previously impossible. Machine learning will have a major impact on robotic capabilities and will likely become a fixture in all robotic systems one day.
This article originally appeared on the Robotics Online Blog. Robotic Industries Association (RIA) is a part of the Association for Advancing Automation (A3), a CFE Media content partner. Edited by Chris Vavra, production editor, Control Engineering, CFE Media, firstname.lastname@example.org.