A new deep learning software enables sorting robots to respond faster and grasp objects more smoothly. According to the developers from the University of California, Berkeley (UC Berkeley), this will allow e-commerce warehouses to automate certain processes and assist human workers in assembly tasks. The goal of the project was to accelerate the “computing time of the motion planner” in robots. In other words: to make the motions both quicker and smoother. To achieve this, the researchers created a deep learning neural network to train the robots.
Principal researchers Ken Goldberg, a William S. Floyd Jr. Distinguished Chair in Engineering at UC Berkeley, and Jeffrey Ichnowski, a postdoctoral researcher at UC Berkeley, both explored in an initial study how to compute both the motion of the robot picking up an object and how it moves when it transfers the object. They called this process a Grasp-Optimized Motion Planner, which was able to successfully compute both motions in approximately half a minute. Unfortunately, the process still wasn’t entirely perfect and made the robots movements too jerky.
The current study managed to enhance the Motion Planner by using a deep learning neural network that resulted in a much faster computing time. Likewise, the researchers also discovered that the Motion Planner could optimize the neural network’s approximation.
Neural networks are designed for robots to learn by example and to train them to generalize certain objects and movements. The combination of both the neural network and the Motion Planner managed to cut the average computational time, allowing the robots to more accurately identify objects and move them around to different locations. According to the researchers, the original computation time of 29 seconds was reduced to as little as 80 milliseconds—less than one-tenth of a second.
This kind of technology system can be extremely beneficial for warehouses and manufacturers that perform sorting tasks. As online shopping continues to grow amid the COVID-19 pandemic, warehouses are experiencing logistical challenges in trying to meet the massive demand. These facilities still tend to rely heavily on human workers, primarily because certain sorting tasks require very nuanced motions and judgment due to the variety of products being managed and transferred.
“Warehouses are still operated primarily by humans because it’s still very hard for robots to reliably grasp many different objects,” explained Goldberg. “In an automobile assembly line, the same motion is repeated over and over again, so that it can be automated. But in a warehouse, every order is different.”
In addition, warehouses generally run on more time-restricted conditions, meaning that a minute delay could be critical to the entire operation. This neural network presents an opportunity to improve the speed and accuracy of robots and allow them to support human workers in the future.
Goldberg has expressed that neural networks, coupled with current advancements in robotic technology, could soon see more robots in warehouses, especially with the boom in online shopping.
“Shopping for groceries, pharmaceuticals, clothing, and many other things have changed as a result of COVID-19, and people are probably going to continue shopping this way even after the pandemic is over,” shared Goldberg.
The study was published in Science Robotics on November 18 and can be accessed here.
Share with us in the comments what you think of an increased use of robots in warehouse environments.
For more news and stories, check out how machine learning is helping in the fight against COVID-19 here.